Drug research and development (R&D) plays a crucial role in supporting public health. However, the traditional drug-discovery paradigm is hindered by significant drawbacks, including high costs, lengthy development timelines, high failure rates, and limited output of new drugs. Recent advances in micro/nanotechnology, along with progress in computer science, have positioned microfluidics and artificial intelligence (AI) as promising transformative tools for drug development. Microfluidics offers miniaturized, multiplexed, and versatile platforms for high-dimensional data acquisition, while AI enables the rapid processing of complex, large-scale microfluidic data; together, they are accelerating a paradigm shift in the drug-discovery process. This paper first outlines the mainstream microfluidic strategies and AI models used in drug R&D. It then summarizes and discusses real-world applications of the integrated use of these technologies across various stages of drug discovery, including early drug discovery, drug screening, drug evaluation, drug manufacturing, and drug delivery systems. Finally, the paper examines the main limitations of microfluidics and AI in drug R&D and offers an outlook on the future convergence of these technologies.
Du Qiao, Hongxia Li, Xue Zhang, Xuhui Chen, Jiang Zhang, Jianan Zou, Danyang Zhao, Weiping Zhu, Xuhong Qian, Honglin Li.
The Convergence of Artificial Intelligence and Microfluidics in Drug Research and Development.
Engineering, 2025, 55(12): 125-174 DOI:10.1016/j.eng.2025.07.025
Drug research and development (R&D) is one of the most challenging and complex areas of the pharmaceutical industry, encompassing multiple stages from initial discovery to clinical trials [1,2]. Innovative drug-discovery paradigms are essential for addressing emerging diseases, overcoming health challenges, and improving public health. Over the past two decades, technologies such as high-throughput screening, animal models, in vitro biology, and structural biology have significantly advanced drug discovery. Despite these developments, persistent issues—including high costs, lengthy development timelines, poor predictive accuracy, disease heterogeneity, and ethical concerns—continue to make the discovery and launch of new medicines both expensive and time-consuming [[3], [4], [5]]. To reduce the cycle time and cost of drug development, we must shift our thinking about the process.
Microfluidics offers an advanced platform for miniaturized, automated, and parallelized analysis, uniquely integrating bioengineering, multimodal sensors, high-resolution imaging, and precise fluid manipulation [[6], [7], [8]]. Its application in drug R&D is transforming the field by providing innovative tools for early drug discovery, preclinical studies, clinical trials, and post-marketing surveillance. The integration of microfluidic systems enables single-cell genomic analysis on compact experimental platforms, opening new avenues for target identification and validation [9]. Owing to its flexibility and scalability, microfluidics is well-suited for various drug-screening approaches. It can accommodate different scales and contexts, including phenotypic screening, diverse compound libraries, a range of cell lines, and combinatorial screening strategies [[10], [11], [12], [13]]. Additionally, microfluidics supports the development of patient-specific organ and tissue models, advancing the field of precision medicine by enabling treatments tailored to individual patient characteristics [[14], [15], [16]]. Its application in drug manufacturing also offers significant advantages, such as high heat- and mass-transfer efficiency, accelerated reaction rates, and precise control over product quality [[17], [18], [19]]. However, the very features that make microfluidics powerful—namely, its multiplexing, versatility, and high-throughput capabilities—also result in the generation of complex, large-scale data that is challenging to analyze manually.
Artificial intelligence (AI) is capable of processing both structured and unstructured data, allowing it to efficiently mine the vast datasets generated by microfluidic systems and rapidly extract meaningful insights [[20], [21], [22]]. The demand for high-quality data to train AI models aligns perfectly with the capabilities of microfluidics. To date, AI has been applied across all stages of drug discovery, benefiting from the rich, large-scale data provided by microfluidic platforms. This integration has significantly transformed the drug R&D paradigm [[23], [24], [25], [26], [27]]. A key function of AI is analyzing large-scale biological and pharmacological data to identify and predict potential patterns. Forward prediction and reverse engineering in the synthesis of nanopharmaceuticals are notable applications of this technology [28,29]. The extraction of morphological features from high-throughput imaging is another important application of AI [30]. In addition, advances in AI have proven valuable in pharmacokinetic/pharmacodynamic (PK/PD) modeling, optimizing dosing regimens, and generating novel chemical structures [31,32]. Conversely, microfluidics can leverage AI capabilities in intelligent decision-making and adaptive feedback mechanisms, shifting the drug development process from mere automation toward true intelligence.
By focusing on the integration of AI with microfluidics, this review aims to provide a comprehensive perspective on how their convergence is creating a transformative tool that is driving a paradigm shift in drug discovery. The closed-loop interaction between microfluidics and AI—spanning data generation to decision optimization—is enabled by multidimensional synergy and dynamic feedback mechanisms. As illustrated in Fig. 1, we outline the synergy between AI and microfluidics throughout the entire drug R&D process. First, we briefly review the fundamental principles of microfluidics and AI in the context of drug development to offer theoretical guidance for stakeholders. Next, we present their applications across key stages of drug R&D, including early drug discovery, drug screening, drug evaluation, drug manufacturing, and drug delivery. Finally, we discuss the challenges associated with the integration of AI and microfluidics in drug R&D and explore future trends in their convergence. We hope this review will inspire practitioners in drug discovery by elucidating essential principles and technologies, thereby accelerating the ongoing paradigm shift in the field.
2. Fundamentals of microfluidics and AI
Since its development in the 1990s, microfluidics has expanded from what was initially referred to as the “micro-total analysis system (μ-TAS)” [33,34] to continuous flow media, discrete flow media, and bionic microfluidic platforms mimicking the human microphysiological system (MPS). The advantages of microfluidics—which include high throughput, low reagent consumption, multiplexing, and multi-module integration—make it a powerful data source for building and implementing AI [7] by addressing the most laborious but important data collection and preparation tasks [8]. In this section, we first introduce mainstream microfluidic systems, including continuous microfluidics, droplet microfluidics, and organs/organoids-on-chips, and outline how these microfluidic platforms can be used for drug discovery. We then discuss the AI models, data acquisition, and processing used to enhance microfluidic drug discovery and how AI can be deployed in this process.
2.1. Fundamentals of microfluidics
Microfluidics is a cutting-edge discipline focused on the precise manipulation of ultra-small fluid volumes (10-9 to 10-18 L) within microscale channels. Surface forces dominate over bulk forces at this scale, leading to increased material transfer, heat transfer, and reaction efficiency. Microfluidics is valuable at all stages of drug discovery, offering innovative opportunities in drug synthesis, screening, and preclinical testing [17]. Here, we explore the principles, control conditions, and applications of continuous-flow microfluidics, droplet-based microfluidics, and organs/organoids-on-chips techniques, classified by their microfluidic technologies.
2.1.1. Continuous-flow microfluidics
Continuous-flow microfluidics is an approach for efficient multi-component fluid mixing, precise flow control, and process enhancement through microscale channels. It is characterized by flow that is continuous and uninterrupted, with no apparent breaks or separations and with mutual solubility between the component fluids. Due to its excellent diffusion mixing controllability and tunable device design, continuous-flow microfluidics has emerged as an alternative to traditional macro-pilot-scale continuous-flow processes in pharmaceuticals [27]. The predictable fluid flow behavior in laminar regimes allows for precise control of drug release rates and spatial distribution within the microfluidic device, as demonstrated in the development of nanodrug carriers [35], lipid-based nanodrug delivery systems [[36], [37], [38]], and customized drug formulations [39].
Continuous microfluidics consists of two main forms: diffusion-based mixing and convection-based mixing. Diffusion-based microfluidics achieves controlled inter-component mixing at nanoscale distances through hydrodynamic focusing [[40], [41], [42], [43]]. In this process, a central flow carrying solvent lipids is focused by an external sheath flow into a narrow region, as shown in Fig. 2(a) [40]. Here, self-assembled monolayers of lipids with remarkably uniform size are produced through the controlled free diffusion of molecules [43]. The thickness of the lipid layer can be controlled by adjusting the ratio of flow velocities among the components [35,44]. Convection-based continuous microfluidics employs specialized microstructures or channel geometries [[45], [46], [47]] and introduces secondary flows into laminar systems to enhance convection and accelerate molecular momentum and mass transfer, as shown in Fig. 2(b) [48]. This approach compensates for the limitations of diffusion-based methods. Microstructure-induced nonlinear hydrodynamic effects are influenced by the Reynolds number, Re, and the geometrical parameters of the microstructure [43,44], as follows:
$R e=\frac{\rho u D_{\mathrm{h}}}{\eta}$
where ρ represents the fluid density, u denotes the fluid velocity, Dh is the microchannel feature size, and η is the dynamic viscosity. Thus, this approach conveniently improves the production process by optimizing the design of microstructures and flow conditions, making the fabrication process of continuous flow microfluidic devices simpler and more robust [49,50].
In continuous-flow microfluidics, the self-assembly of nanoparticles is driven by rapid mixing-induced nanoprecipitation and temporal solubility gradients [51]. Therefore, precise control of the ratio (μ = Tmix/Tagg) of mixing time (Tmix) to nanoparticle nucleation time (Tagg) is a critical process condition that determines the size and monodispersity of the nanoparticles [19,52]. Small and well-monodispersed nanoparticles are easily obtained when μ < 1, while large-sized and polydispersed nanoparticles are more likely to be formed when μ > 1. Since the system is completely in the laminar flow region and the microscale confined space provides extremely short mass-transfer distances for molecular transfer, the microfluidic-based mixing time can be reduced to 10-4 s, which is a 103-fold increase in mixing efficiency compared with macroscopic mixing. Mass-transfer efficiency is further enhanced by the chaotic advection [53]. These spatiotemporal cross-scale features enable continuous-flow microfluidic systems to precisely control the ordered motion of drug molecules, resulting in mild reaction conditions and higher product purity and quality. The Péclet number (Pe) characterizes the ratio of convection rate to diffusion rate in a continuous-flow microfluidic system:
where λB represents the diffusion coefficient, and Tconv and Tdiff respectively denote the convection and diffusion times. Higher Pe indicates that convective transport is more significant than diffusive transport, a situation that is usually associated with faster mixing efficiencies and more homogeneous mixing results [54]. The ratio between the viscosity coefficient and the diffusion coefficient is characterized by the Schmidt number (Sc):
where ν denotes the kinematic viscosity coefficient. Sc can be linked to the mixing time, the chemical reactions that occur along with the transport of substances [55], and the laminar residence time distribution [56]. The relative timescale between chemical reactions and fluid flow in a micromixed system with chemical reactions is described by the Damköhler number (Da):
$D a=\frac{T_{\mathrm{rea}}}{T_{\mathrm{conv}}}$
where Trea represents the one-step chemical reaction time. Da has been employed to evaluate the relative rapidity of a chemical reaction with respect to the convective mass transfer and to the mixing reaction yield [57]. The fundamental physicochemical processes in continuous-flow microfluidic systems are well understood using these dimensionless numbers.
Continuous microfluidics makes it easy to design microfluidic channels and adjust the flow rate to achieve large-scale, low-cost concentration gradient arrays with high spatial and temporal resolution, allowing drug R&D to greatly benefit from this precise platform [[58], [59], [60]]. This high-throughput approach enables the simultaneous testing of multiple drugs or compounds at different concentrations on a single device, resulting in a large amount of information being obtained at the same time and thereby providing low-cost and high-content data for AI model construction [23,61]. The incorporation of active substances such as cells and biomarkers into the system also allows for the determination of drug uptake, distribution, and metabolism and toxicity assessment by monitoring cell growth, apoptosis, protein expression, or the activity of biomolecules [[62], [63], [64]]. The optimal working concentration range of a drug and the effect of a drug combination can be determined by optimally designing a concentration gradient network to assess the synergistic effect of a multi-drug combination, which is important for adjusting drug regimens, optimizing drug dosages, and reducing side effects [65]. Most importantly, this method enables the modeling of time-dependent drug effects [66], allowing for a detailed understanding of how drug concentration correlates with efficacy in pharmacokinetics (PK). This capability is crucial for assessing pharmacodynamics (PD) and safety over various time scales, including cumulative, delayed, and emergent drug effects [14,67,68].
2.1.2. Droplet-based microfluidics
Droplet-based microfluidics is a key area within microfluidics that enables the creation and control of numerous separated microdroplets within microchannels with precise geometries. These microdroplets have interfaces that do not mix with the carrier fluid [69] and a significant pressure gradient. Droplet microfluidics offers a versatile approach for creating microchambers of various sizes, compositions, and shapes, including Janus [70], core-shell [71], multi-compartment [72], and other heterogeneous particle structures [[73], [74], [75], [76], [77]]. These droplets with a specific size distribution and structural composition offer extraordinary advantages. First, they act as highly efficient reservoirs for biochemical reactions. Their tiny volumes, ranging from femtoliters to nanoliters, allow for rapid heat and mass transfer over short distances, significantly boosting mixing efficiency and reaction rates while preventing cross-contamination between samples [54]. Second, microdroplets with well-defined structures and components can be used as templates for the synthesis of pharmaceutical microspheres and microcapsules [17]. Third, the droplets allow for parallel scaling, providing an efficient platform for large-scale analysis for high-throughput biomedical analysis and drug screening [78,79].
Droplet generation is the first step in droplet-based microfluidics. The microfluidic geometries commonly used for droplet generation include T-junction, flow-focusing, and co-flow [69,80], as shown in Figs. 2(c)-(e). Although there are many different geometries for droplet generation, the underlying mechanisms all involve a balance of competing forces at the phase interfaces [81]. The competition between interfacial tension and viscous forces is characterized by the capillary number (Ca):
$C a=\frac{u \eta}{\gamma}$
where γ denotes the interfacial tension. A large Ca indicates that viscous forces dominate, leading to bending deformation of the fluid interface and resulting in small, high-frequency droplets. Conversely, a small Ca suggests that surface tension is the primary force, making the interface less prone to bending deformation and favoring the formation of larger, low-frequency droplets. The Ca in microfluidization is typically in the range of 10-4-10 [69]. As the characteristic velocity increases, the role of fluid inertia is no longer negligible, and the competition between inertial forces and surface tension can be expressed in terms of the Weber number (We):
$W e=\frac{\rho u^{2} D_{\mathrm{h}}}{\gamma}$
In co-flow, when the inner phase is fast and the outer phase is slow, the viscous force of the outer phase is minimal. As a result, the breakup of the droplet is primarily influenced by the inertial force of the inner phase [82]. In addition, other dimensionless numbers including viscosity ratios, flow-rate ratios, and geometric feature size ratios can be used to tune the droplet generation characteristics [80,83].
Droplet microfluidics is a promising technology that facilitates parallelization for high-throughput drug synthesis, bridging the scale from lab-scale to industrial-scale applications for drug development [78,84]. To date, tree and trapezoidal networks are the two most commonly used architectures for parallelized droplet-based microfluidic design. In the tree network architecture, each channel splits into two subchannels. This division then continues, resulting in the distribution of the flow across a total of 2N channels, which forms a tree structure with N layers of branches [[85], [86], [87]]. The trapezoidal network architecture takes a parallelization approach in which the droplet cells are arranged in a series of rows and columns in a matrix fashion [[88], [89], [90], [91]]. Droplet cells are connected in parallel, so clogging does not cause pressure fluctuations in neighboring cells [84]. The trapezoidal network architecture allows the droplet generation unit integration to reach approximately 103 cm−2 and the throughput to reach 1.5 L∙h−1, while the coefficient of variation can be controlled at < 5% [89]. Microdroplets with complex components can also be prepared using parallelized droplet microfluidic devices [92,93].
The functionalization of droplets is crucial for enabling their use in drug discovery. Such manipulation can be divided into confined droplet microfluidics and open-droplet microfluidics, based on whether the microdroplets are confined by walls. Confined droplet microfluidics allows for complex functionalization within a closed channel, while open-droplet microfluidics enables functionalization in a highly controlled manner on an external surface using various external field forces, such as acoustic, optical, electrical, and thermal forces [83,94]. Regardless of the type of manipulation, microdroplets facilitate the automation of process such as splitting, merging, selecting, mixing, and extracting, which are important at every stage of the drug-discovery cycle [[95], [96], [97]]. Different drug compounds or reaction conditions can be encapsulated within droplets for the rapid, high-throughput evaluation of multiple compounds. By mixing drug droplets with varying concentrations, researchers can control the drug concentration gradient in a precise and programmable manner. This method is useful for drug dose-response testing, as it helps determine the optimal therapeutic dose [98]. Additionally, mixing enhanced by active or passive oscillatory flow improves liquid-liquid extraction and sample analysis. This technique aids in exploring new synthetic pathways or understanding drug metabolic processes [99]. Furthermore, encapsulating single cells, proteins, nucleic acids, exosomes, and other biomolecules within microdroplets supports high-throughput drug target selection, lead identification, and preclinical studies [100].
Droplet-based microfluidics holds great promise for drug synthesis, drug screening, and other stages of drug R&D. It introduces a new paradigm for understanding drug-development principles and provides a technical foundation for large-scale production. Moreover, precisely functionalized droplets have significantly advanced drug R&D by serving as a platform for high-throughput drug analysis. The ease of scalability facilitates widespread application and enables the delivery of high-content microfluidic information, including size, shape, structure, components, and biochemical cues. This generates a database for parallel analysis in large-scale drug testing, improving the learning and analytical capabilities of AI [101].
2.1.3. Organs/organoids-on-chips
Organs-on-chips (OoCs) are advanced MPSs that employ microfluidics to replicate some or all of the physiological functions of tissues and organs. By continuously perfusing and culturing human cells within microscale chambers, OoCs provide precise control over cellular structures, tissue hierarchies, physicochemical environments, and vascular networks [102,103]. OoCs have shown partial potential to replace traditional preclinical models such as animal models and two-dimensional (2D) cell cultures. They also enable the reconstruction of complex organ-level physiological functions and clinically relevant microenvironments [104]. Organoids are three-dimensional (3D) in vitro organ models formed from self-organizing, stem-cell-derived multicellular structures. Organoid microarrays can replicate organ function and microstructure, simulate disease mimicry, and support long-term culture. Due to these significant advantages, organoid microarrays have been widely used in the study of various types of diseases, including tumors, and in drug screening and evaluation [105].
The first step in developing prototypes of engineered OoCs is material selection. The design constraints and patterning capabilities are dictated by the machinability of the chosen material. Biocompatible and flexible materials are preferred, as they can support cell culture and mimic the mechanical cues of native human tissues. Polydimethylsiloxane (PDMS) is a widely used material in polymer microfabrication due to its compatibility with traditional soft lithography processes, its excellent machinability, and its capability for precise patterning. With a Young’s modulus of 1-3 MPa, PDMS is highly pliable and super-elastic, allowing it to replicate a wide range of tissue mechanical properties, including the mechanical stimulation of cardiac tissue and alveolar cyclic stretching, as shown in Figs. 2(f) and (g) [106,107]. In addition, PDMS has good light transmission, which enables optical microscopy to visualize fluid flow and morphology within the system. However, PDMS readily absorbs hydrophobic small-molecule drugs, potentially compromising experimental accuracy [108]. Consequently, chemically inert polystyrene has become a standard material for drug-related applications due to its optical transparency, good processability, and biocompatibility [109]. In contrast to PDMS, polystyrene is a rigid thermoplastic. To overcome this limitation, it is often combined with other low-absorption flexible materials, such as hydrogels, polyurethane elastomers, styrenic block copolymers, and polycarbonate-thermoplastic elastomer hybrids [[110], [111], [112], [113]].
Effective operation of microphysiological models relies heavily on the reasonable control of flow, mechanical, biological, and chemical constraints on the OoC. Establishing human physiological fluid dynamics is the foundation for organ morphogenesis and simulating pathophysiological environments. Fluid flow triggers biologically active processes such as cellular accretion, orientation, differentiation, and migration by inducing shear and compressive stresses [114]. Moreover, on-chip flow patterns—including oscillation, pulsation, turbulence, and air-liquid exchange—are crucial for replicating the basic physiological flow environment of organs, such as small intestinal villous crypts, vascular wall shear, and alveolar interfacial material exchange [[115], [116], [117]]. By inducing mechanical forces at tissue flow-solid interfaces and within cell-culture matrices, organoids improve our understanding of the mechanical cues influencing protein expression and adverse effects on the efficacy of inflammatory drugs [118,119]. Modeling organ-level physiological microenvironments necessitates independent control of vascularization and multiple tissue types that constitute hierarchical structures. This control provides the biological basis for OoCs to reproduce the physiological functions of human organs [120]. The creation of biochemical microenvironments comprising cells, tissues, and organs can lead to the growth, differentiation, metabolism, and functional expression of multilevel aggregates. These micro-engineered models replicate the complex interactions between different cell types through cytokines, hormones, gaseous environments, pH, ionic equilibrium, and metabolites. They play a critical role in predicting and evaluating drug metabolism, bioavailability, and drug-cell interactions [121].
Finally, it is essential to ensure the fidelity of the cell or tissue culture on the chip. The first step in constructing organoids and OoCs involves selecting an appropriate cell source and type [122]. Real organs typically comprise multiple cell types, each serving a specific function. For example, the lungs consist of epithelial cells, endothelial cells, and immune cells [45]. In OoCs, synthesizing these different cell types is crucial for maintaining the physiological relevance of the model. Primary cells, which are isolated directly from tissues, tend to remain closer to their physiological state and function in vivo. However, with successive cell passages, issues such as phenotypic drift, altered gene expression, and decreased functionality may arise, leading to less reproducible models [123]. Induced pluripotent stem cells (iPSCs) process the unique ability to self-renew and differentiate into diverse cell types, making them ideally suited for constructing various organ models [124]. Under specific culture conditions, stem cells can be induced to differentiate into specific cells of the target organ, such as cardiomyocytes [125] and neurons [126]. Three-dimensional cellular construction also facilitates the formation of intercellular interactions and structural arrangements that allow for complex cellular activities, including cell adhesion, signaling, and cell migration [127]. Furthermore, 3D cell culture enhances physiological functions (e.g., metabolism, secretion, and response) and promotes the differentiation of stem cells into specific cell types, which leads to better reconstruction of organ function.
A consideration of cell behavior is critical in developing organoids that mimic real structures and functions. One important factor is cell polarity, which refers to the asymmetry cells exhibit in form and function, often resulting in functional differences across various regions of the cell membrane. Maintaining cell polarity within organoids is vital for reproducing the true functions of these organs. For example, the polarity of intestinal epithelial cells influences nutrient absorption and secretion [128]. Cell signaling, the process of cellular communication through molecular signals, is the fundamental mechanism enabling organoids and OoCs to recapitulate tissue physiology and respond to stimuli [129]. Therefore, faithfully mimicking cell signaling networks is critical for organoids and OoCs to accurately model human biology and disease.
OoCs can be seamlessly integrated into the drug-discovery process [130,131]. In the early stages of drug development, these MPSs serve as effective tools for disease modeling, target identification, and lead optimization. During the preclinical screening phase, a hepatocyte MPS was shown to predict the absorption and excretion of drug compounds, closely aligning with the results from clinical studies [132,133]. OoCs offer a reliable and cost-effective means of predicting the efficacy of drug compounds in the human body, effectively addressing the clinical biases associated with traditional animal models and 2D culture systems. PD modeling based on MPSs has been applied in tests evaluating effective drug concentrations for the liver and for conditions such as breast cancer and uterine cancer [134,135]. Ultimately, OoCs have the potential to predict the in vivo efficacy of a drug and the anticipated toxic response in humans during preclinical trials. For example, a liver-on-a-chip facilitates the assessment of drug toxicity and metabolism [136,137]. Lung cancer organoids have shown promise in drug safety assessment and improving clinical trial success [138]. Additionally, by employing fluidic networks to interconnect organoids in a physiologically relevant sequence, researchers can create a human microchip that enables a comprehensive evaluation of drug delivery, absorption, metabolism, and circulation [105,139]. Platforms that utilize patient-derived stem cell OoCs are also playing a crucial role in later stages of clinical development, where drugs are initially tested in vitro before in vivo administration. Ultimately, this approach has potential to advance personalized medicine platforms [15].
OoCs and organoids present several advantages over traditional animal models: They offer a more realistic in vitro simulation of the human physiological environment while also addressing ethical concerns and eliminating species differences. Furthermore, these MPSs enable the development of personalized organoids from patient-derived materials. Such systems can more accurately predict individual drug responses and provide critical data for precision medicine. AI models leverage this data to identify drug patterns and understand the mechanisms of action in various physiological conditions. Additionally, AI can process large-scale data automatically, providing immediate feedback and optimization suggestions, which significantly accelerate the drug-development cycle. The AI algorithms that enhance microfluidic-based drug discovery and development are discussed in detail in the next section.
2.2. Fundamentals of AI
AI has experienced a boom as the science used to simulate human consciousness and thought is applied in R&D. Owing to the development of hardware and software technology and the advent of the big data era, AI has given birth to advanced algorithmic frameworks such as machine learning (ML) and deep learning (DL). In particular, the recent convergence of AI and drug design has proved the potential of AI in the field of drug development [24]. With the rapid growth of microfluidic data and the development of high-throughput sequencing technologies, AI shows great potential for addressing the challenges associated with high heterogeneity and multidimensional omics data analysis. In this section, we focus on the use of AI for various stages of microfluidic drug R&D, with an emphasis on ML and DL models and how they can benefit microfluidic-based drug R&D, as shown in Fig. 3. The data-acquisition and processing processes used for training are crucial and are therefore also discussed. Finally, how AI models are deployed in microfluidic-based drug R&D is discussed in detail. The concepts, applicability scenarios, and shortcomings of various AI algorithms are briefly described, but excessive technical details are not provided.
2.2.1. Machine learning
ML is a branch of AI that essentially uses algorithms to learn data features and make judgments and predictions about the future state of new datasets. ML includes supervised learning, unsupervised learning, and reinforcement learning (RL). In supervised learning, a model is trained using a labeled dataset, which allows the model to then make predictions about new examples based on what it has learned. Supervised learning is applicable to classification and regression problems. At present, widely used supervised learning algorithms include the k-nearest neighbor algorithm (KNN), gradient boosting machine (GBM), random forest (RF), decision trees (DTs), support vector machines (SVMs), linear regression, logistic regression (LR), and neural networks (NNs). The results obtained through supervised learning algorithms are based on probabilistic statistics and are strongly influenced by feature selection and quality.
Unsupervised learning complements supervised learning by enabling pattern recognition and the prediction of samples using unlabeled datasets. In practical problems, there is insufficient prior knowledge to categorize the data, and manual labeling can be prohibitively expensive. Therefore, unsupervised learning is primarily employed to track clustering and dimensionality reduction problems in datasets without labeling information. This approach helps uncover hidden patterns in the input data and organize the data into meaningful clusters. Principal component analysis (PCA), Gaussian mixture, k-means clustering, hierarchical clustering, isolation forest, autoencoder, and independent component analysis (ICA) are the most commonly used unsupervised learning algorithms. However, the unpredictability of unsupervised learning results and the difficulty of quantifying the accuracy of the results pose significant challenges.
RL consists of an intelligent body and an environment. The intelligent body takes various actions in the environment and receives rewards or punishments from the environment as feedback. The intelligent agent selects payoff-maximizing strategies to achieve a specific goal through a large amount of feedback. Monte Carlo, temporal difference, proximal policy optimization, deep deterministic policy gradient, adaptive dynamic programming, Q-learning, and deep Q-networks (DQNs) are widely used RL algorithms. RL strategies are particularly effective for developing autonomous decision-making systems, such as droplet-based microfluidics for path optimization [140]. These ML algorithms are well-suited to handle the large datasets generated by microfluidics and have been widely used in areas such as circulating exosomal biomarker analysis [141], heterogeneous biological systems stratification [142], and sepsis diagnostics [143].
Different ML algorithms are suited to various types of drug development processes. Choosing appropriate models for specific problems is crucial in order to utilize the data-mining capabilities of the algorithms and uncover potential implicit interactions within the data. A supervised learning model was used to predict outcomes in patients with four drug-based diffuse large B-cell lymphomas (DLBCL), identifying rational targets for intervention [144]. Clustering through unsupervised learning, followed by classification using supervised learning, was used to identify drug-resistant phenotypes in the cytotoxic treatment of pancreatic cancer [145]. Unsupervised learning was used to facilitate an evaluation of the effects of drug concentration and duration of action on U937 human leukemia cells on a polymethyl methacrylate (PMMA) chip with an indium tin oxide (ITO) substrate [146]. An RF model was utilized to construct classifiers to assess the neurotoxic effects of 6-hydroxydopamine (6-OHDA) on neurons in brain-like organs. It was also able to predict highly intricate images and distinguish between various treatment conditions [147]. ML offers significant advantages in predicting the physicochemical properties of drug crystals in the context of drug crystallization [148]. Bayesian optimization plays an important role in determining conditions such as the flow-rate ratio, buffer pH, and total flow rate in lipid nanoparticle (LNP) preparation [149]. Knowledge-based computational disease modeling is a valuable complementary tool that makes it possible to generate images of the overall mechanisms of disease in order to best simulate in vivo disease conditions [150].
2.2.2. Deep learning
DL, which represents the cutting edge of ML, has particular applicability in processing microfluidics to generate complex data patterns and large-scale datasets [151]. It focuses on learning high-dimensional abstract features of data through a multilayered NN structure. A neuron is the basic unit of an NN and can be considered as a node, with each neuron receiving input from the previous layer of neurons and producing output. Each neuron is usually associated with an activation function that introduces nonlinear features. In terms of overall architecture, an NN comprises several layers, including an input layer, several hidden layers, and an output layer. The output layer can have many nodes, with each output node corresponding to a specific task to be predicted. Forward propagation and back propagation are two core computational processes in NNs. In the forward propagation process, the input data is transmitted through the input layer to the hidden and output layers, where each neuron weights the input according to the weights and biases and passes the result to the next layer. This process does not involve updating the model parameters; rather, the input data is merely passed through the network and the predicted values are calculated. Back propagation is a key training and optimization step in DL. It reduces the difference (loss) between the predicted and true values by calculating the gradient of the loss function with respect to the model parameters (weights and biases). The weights and biases are updated sequentially from the output layer to the input layer. This process employs the chain rule for gradient calculation, which allows the model parameters to be adjusted incrementally, thereby bringing the model’s predicted values closer to the real values [1].
DL offers the advantage of various flexible NN architectures, each with specific structures that are useful in the various stages of drug discovery and development. Feedforward NNs consist of multiple layers of neurons without any feedback connections among the neurons in each layer. Therefore, data can only flow unidirectionally in the input-to-output direction. This architecture is the simplest DL architecture, represented by the multilayer perceptron (MLP). A convolutional neural network (CNN) is a type of feedforward NN that comprises a convolutional layer and a pooling layer. The convolutional layer is designed to extract spatial structural features from the input data, while the pooling layer serves to reduce the dimensionality of the feature map. CNNs are extremely advantageous in processing image and video data to effectively capture spatial hierarchical features. For example, CNNs have been used to analyze molecular image processing and drug cell response [8].
Another structure is the recurrent neural network (RNN), which uses a chain of repeating modules to process sequential data. In each time step, the RNN receives the input of the current time step and the hidden state of the previous time step, and outputs the hidden state of the current time step after processing through the activation function. This makes it possible to capture temporal dependencies in data in scenarios where persistent information is required. This approach is suitable for processing text, speech, and time series data, for example. Bidirectional recurrent neural networks (BRNNs), deep recurrent neural networks (DRNNs), and long- and short-term memory networks (LSTMs) are commonly used RNNs. Among them, LSTMs are used to process the biologically active sequences of drugs in order to help establish and predict drug-biomolecule interactions [152].
Finally, generative adversarial networks (GANs) are network architectures consisting of two main NN models: a generator and a discriminator. The generator produces an output similar to the target data from random noise and other input data, with the goal of tricking the discriminator into being unable to distinguish the generated data from the real data. The discriminator is a binary classifier designed to distinguish between the data generated by the generator and the real data. It takes inputs and outputs a probability value from 0 to 1 indicating the probability that the input data is the real data. An LSTM-based generative DL model was used as a “chemical language model” to generate hypothetical liver X receptor alpha agonists [152]. Generative models also make it possible to predict potential safety issues with generated molecules by comparing them with known toxicity data. Popular generative models in DL include variational autoencoders (VAEs) and stream generation models, which allow controllable generation of the expected macromolecules in the ab initio design of drugs [153].
There are several issues to keep in mind when deploying DL models in microfluidic-based drug discovery and development. First, the quality and quantity of the data determine how well the DL model is trained and how well it performs. It is critical to ensure the accuracy, completeness, and representativeness of the dataset. Data processing is more than 80% important for the successful implementation of DL [1]. A detailed discussion on data acquisition and processing is provided in the next section.
Second, overfitting and underfitting phenomena are common problems in the model learning and prediction process. In overfitting, a model performs well in the training set but generalizes poorly; in underfitting, a model is unable to adequately capture key features and patterns in the data. The cause of overfitting is that the model is so complex that the noise in training is also learned. Therefore, it is possible to avoid overfitting through dataset augmentation, regularization, feature selection, and early stopping. Underfitting is often caused by models that are too simple. Therefore, it can be solved by increasing the model complexity, adding features, decreasing regularization, and adjusting hyper-parameters.
Third, DL models typically require significant computational resources. Effective and rational utilization of hardware gas pedals such as graphics processing units (GPUs) and tensor processing units (TPUs), as well as the allocation of cloud computing resources, are important considerations for model cost and performance tradeoffs. Finally, it is worth noting that DL has been referred to as a “black box” model, which means that its decision-making process is sometimes difficult to understand—especially in the field of drug discovery, where a high degree of interpretability and transparency is required [1,154].
2.2.3. Data acquisition and processing
The collection and processing of datasets is usually the first step in building an AI model. In a supervised learning task, the dataset should include a description of the features of the system and the corresponding results produced by these features. Common drug or microfluidic input characterization data, such as shear rate, drug solubility, melting point, mixing rate, or emulsification conditions, can be obtained through computer simulations or experimental measurements. The processing and screening of data is an essential step. Feature engineering is the process of identifying and constructing relevant features that can significantly improve the predictive power of AI models. However, while adding more features may provide richer information, expanding the feature set may also introduce spurious correlations that can affect the predictive accuracy of the model. Therefore, the data must be further processed to improve the predictive accuracy and generalization of the AI model.
A large amount of data is usually used when training an AI model, and obtaining high-quality data is usually the most complex and time-consuming step. The amount, consistency, and standardization of data play an important role in increasing the usability and predictive power of AI algorithms during drug development. At present, in order to cope with the problem of insufficient data, government officials provide large-scale public databases for researchers to use, such as the large-scale datasets managed by the National Center for Biotechnology Information (NCBI) in the United States, which include DNA sequences, RNA sequences, protein structures, and molecular and genetic medicine [155]. Some scholars have also tried to enable an AI model to learn enough from existing data through a small-sample approach [156]. Microfluidics enables the real-time acquisition of data through sensors or imaging techniques [157], which can increase the speed, reproducibility, and data quality of new data generation in an automated and high-throughput manner.
Appropriate processing of the acquired data can improve the performance of an AI model. The data-processing procedure usually includes data cleaning, feature extraction, data normalization, data integration, and data enhancement. In the process of training an AI model, the ML or DL algorithms used for training must be subjected to hyper-parameter tuning, data normalization, regularization, and so forth. In recent years, research on extracting useful features has become increasingly diverse, including PCA, univariate feature selection, recursive feature elimination, and correlation coefficient methods. By removing redundant or irrelevant features, the computational cost can be reduced and the training process accelerated; it also makes the intrinsic structure of the data easier to understand and increases the interpretability of the model. Integrating data from various sources—such as experimental, simulation, and clinical data—into a single unified dataset offers a comprehensive view and minimizes reliance on individual datasets.
2.2.4. Basic processes for applying AI
The application of AI in drug discovery follows several key steps involving data characterization, model building, model and experiment iteration, and result interpretation. A typical workflow can be summarized as “input-AI model-output” [8,158], as shown in Fig. 4. First, the specific drug R&D problem to be solved using microfluidics and AI—such as liposome design or phenotypic screening—is clearly defined, along with the task objectives, constraints, and expected outcomes. In the second step, data collection and preparation is performed. Accurate, complete, and standardized data on microfluidic platforms that is relevant to the objectives is gathered. The data is then integrated and cleaned to eliminate outliers and noise; it may also be manually labeled, if necessary. Third, feature extraction and selection is performed. Relevant features are extracted for each drug and microfluidic experimental condition, utilizing feature sensitivity analysis and correlation analysis to identify the most informative features for subsequent modeling and analysis. In the fourth step, model building and optimization are executed. Appropriate AI models are selected based on specific drug-related problems, and the models are trained using cleaned and selected data. The internal parameters are iteratively adjusted through optimization techniques such as cross-validation and hyper-parameter tuning to increase the predictive and generalization capabilities. The trained model is then used to predict the task goals of interest for future experiments, providing decision support and informing the development of the next experimental strategy based on the model predictions. The entire workflow is optimized and refined through iterative feedback, experimental insights, and data updates to improve the model’s predictive accuracy. In the fifth step, the final prediction results are interpreted. The predicted results are analyzed to explain the drug response in different microfluidic environments in order to reveal potential pharmacological and biological mechanisms.
These steps and the role of AI in enhancing microfluidic-based drug discovery and development have been demonstrated through complete process engineering and hold great potential to drive innovation at all stages of microfluidic-based drug R&D [159].
3. Synergy between AI and microfluidics in early drug discovery
Early drug discovery is the initial stage in the preclinical drug development process aimed at understanding the biological and pathophysiological mechanisms of a specific disease, identifying potential drug targets, and discovering efficacious compounds. However, discovering effective drugs among enormous numbers of drug candidates is time-consuming and laborious. Here, the synergy of microfluidics and AI plays a crucial role. Microfluidics enables the delivery of biochemical information with low cost and high throughput. Microfluidic-based OoCs also permit the faithful mimicry of human systems to identify drug targets with repeatable and predictable results. AI helps in target identification, validation, and the lead compound optimization process, thereby improving the efficiency and success rate of new drug development.
3.1. Target identification
A drug target is the molecule that a drug interacts with through specific binding or regulation within an organism. This target is usually a protein, enzyme, receptor, or other biomolecule within a cell. These targets are crucial in the onset and progression of specific diseases [160]. The final goal of target identification is to discover one or more targets suitable for drug development. Such targets can effectively intervene in the core mechanisms of the disease, facilitating the achievement of a therapeutic effect through pharmacological intervention.
Droplet-based microfluidics serves as a stand-alone microreactor for high-throughput single-cell analysis by encapsulating biological samples, such as cells and molecules, in microdroplets. Omics analysis on microfluidic platforms enables the collection of omics information and high-resolution omics data, including gene expression and proteomic profiles, from samples contained in individual microdroplets [161]. The use of microfluidic microarrays for proteomics sample preparation with data-independent acquisition (DIA) mass spectrometry (MS) as a combined technology pathway has enabled proteomics analysis at the single-cell level. A proteomics chip enables multiplexed and automated cell separation, counting, imaging, and sample processing in a single device, as shown in Fig. 5(a) [162]. By combining microfluidic sample processing with the DIA-MS of specific MS using libraries, approximately 1500 proteomes in 20 individual mammalian cells were analyzed, allowing the identification of important drug targets [162].
Recent studies have shown that exosomes can participate in tumor pathology as carriers of bioactive molecules. Liu et al. [163] proposed a droplet-based immunosorbent assay to detect the expression of target proteins in individual exosomes by measuring the fluorescent signals in droplets. Exosomes were immobilized on magnetic beads through enzyme-linked immunosorbent assay (ELISA) complexes, and the microbeads were then encapsulated into multiple microdroplets, as shown in Fig. 5(b) [163]. In this way, the droplet-based microfluidic platform provides a promising technological pathway for discovering and analyzing potential drug targets based on omics data.
Despite these advantages, the analysis and processing of the complex omics data generated by microfluidics are still challenging tasks, as information on gene expression and cell-specific expression that is difficult to process manually may be hidden in this data. ML enables the evaluation of patterns in microfluidic-based transcription data on a genomic scale. For example, DL frameworks can be applied to perform cluster analysis on single-cell RNA sequencing (scRNA-seq) data, as illustrated in Fig. 5(c) [164]. In this framework, the researchers use iterative smoothing and self-supervised discriminative embedding techniques, which are validated on 17 real scRNA-seq datasets. The algorithm automatically selects positive and negative sample pairs, effectively reducing high-frequency noise and improving data quality [164]. When combined with interpretable AI algorithms, a deep neural network (DNN) also demonstrated the ability to analyze temporal gene expression [165]. A study by Piazza et al. [166] utilized a limited proteolysis (LiP)-based approach to explore histological information in complex eukaryotic proteomes. They developed a supervised-learning-model-based classifier, LiP-Quant, to identify small-molecule compound targets, predict binding sites, and estimate target affinities. In the target-identification task, LiP-Quant can offer a proxy for the drug binding site by identifying peptides that undergo structural changes upon compound binding. It quantitatively assesses the binding specificity of these peptides by estimating the half maximal effective concentration (EC50). Migliozzi et al. [167] developed a computational method called SPINKS to identify master kinases for functional subtypes of human glioblastoma using comprehensive multi-omics data. This probabilistic classification tool can assess the association between treatment response and glioblastoma subtypes, aiding in patient selection for prospective clinical trials. In addition to proteomics-based information, microfluidics generates high-throughput genomics, transcriptomics, and metabolomics data. AI further enhances drug and disease target identification and prediction by leveraging its powerful data-analysis capabilities. Various genomic datasets have been used to construct SVM classifiers, enabling the classification of proteins from breast, pancreatic, and ovarian cancers into drug targets and non-drug targets [168]. While the roles of microfluidics and AI in drug target identification are still largely separate, the potential of AI in processing high-throughput omics data generated by microfluidics has been widely recognized.
OoCs have great potential for early target identification in the drug-discovery pipeline, largely due to their ability to provide close proximity to the actual microenvironment and drug response in vivo. Another advantage of OoC technology is its ability to integrate different types of tissues and organs, thereby providing insights into target availability and interaction networks across these tissues. This integration proves valuable for identifying drug targets in complex biological environments and mechanisms of response to drugs [130,169]. A 3D microfluidic-based model specific for human breast cancer bone metastasis was developed to simulate the complex interactions between circulating metastatic breast cancer cells, vascular endothelium, and bone tissue. This model, used in conjunction with high-resolution real-time microscopy, measured the trans-endothelial migration of metastatic breast cancer cells. The findings suggest that the CXC motif chemokine ligand 5 (CXCL5), produced by osteoblasts, and its receptor CXCR2, expressed by tumor cells, may be potential targets for therapeutic intervention [170]. It is possible that AI could play a similar role in processing OoCs and early drug discovery on disease models. Synergy between OoCs and AI is a powerful tool for revealing potential therapeutic targets and complex mechanisms in vivo [171]. DL is employed to predict new drug targets by processing vast amounts of image data and time-series information generated during the operation of OoC and disease models, as shown in Fig. 5(d) [20,172]. In turn, OoCs can be used to validate potential targets identified by AI models based on genomics and proteomics analysis.
3.2. Drug-target interactions
Diseases can be treated with drugs that intervene in abnormal biological processes through interactions with their targets. Confirming the actual effect of a drug on its target—such as the drug-target binding capacity, specificity, and mechanism of action—is the primary task of target validation. Microfluidic chips are widely used for in situ studies of drug-protein-ligand interactions. These versatile platforms provide an efficient means for studying drug-target interactions, combining techniques such as fluorescent probes (Fig. 6(a)) [173] and ion mobility with natural MS (Fig. 6(b)) [174]. In this way, drug-target affinity—which encompasses binding strength, binding site, interaction patterns, target specificity and selectivity, changes in target structure and function, and the temporal effects of drug-target interactions—can be evaluated effectively. Various types of microfluidic microarrays have been employed to assess the binding strength between cell surface receptors and protein ligands, as well as binding sites, as illustrated in Fig. 6(c) [[175], [176], [177]]. This microfluidic approach permits antigen-specific immobilization on surfaces with controlled hydrodynamics, facilitating the design of more effective immunotherapies. Real-time microfluidic analysis, when combined with surface plasmon resonance imaging, can also be used to study quantitative protein interactions, enabling the screening of anticancer drugs with high specificity and sensitivity, as illustrated in Fig. 6(d) [178]. Additionally, continuous microfluidics has been used for one-particle, one-compound combinatorial peptide library screening in target specificity assessments [179].
DL has demonstrated promising capabilities in dealing with drug-target interaction metrics. The deep bilinear attention network (BAN) framework, DrugBAN, can learn pairwise localized interactions between drugs and targets and adapt to out-of-distribution data, as illustrated in Fig. 6(e) [180]. The predicted results from drug molecular maps and target protein sequences provide interpretable insights. The CNN-based DeepDTA framework allows the prediction of binding affinities for drug-target interactions using only the sequence information of targets and drugs [181]. Predicting sequence-based drug-target interactions requires high input sensitivity, like that of the deep-learning-based ConPLex model, which employs protein-anchored contrast co-embedding to enhance its input sensitivity. This model predicted 19 kinase-drug interactions and experimentally validated 12 interactions, including four sub-nanomolar affinities and a strongly binding EPHB1 inhibitor [182]. In addition, drug-target prediction models—such as DeepAffinity, a semi-supervised DL model based on RNN/CNN; GSL-DTI, which is based on graph structure learning networks; and SILCS-Covalent and BE-DTI, which are based on active learning—have successfully shown potential for affinity prediction [[183], [184], [185]].
Table 1 [[173], [174], [175],[180], [181], [182],184,[186], [187], [188]] illustrates the components of microfluidic integrated monitoring used to evaluate drug-target interactions, as well as the ML models used to process drug targets. It is foreseeable that—by combining microfluidic-based approaches with DL frameworks—the integrated process of microfluidic data collection, AI model construction, and AI prediction of drug-target interactions will facilitate early drug discovery, greatly improve drug-discovery efficiency, and lay the foundation for subsequent lead validation and relevant drug screening.
4. AI for microfluidic high-throughput drug screening
Drug screening identifies potential therapeutic molecules with specific actions on disease targets or physiological systems from a large pool of compounds [189]. It not only evaluates how a drug affects a disease target but also assesses the drug’s impact on the overall biological system, with a focus on toxicity and safety. Identifying potential toxicity issues early in the screening process helps prevent safety problems in later clinical trials and increases the chances of successful development [2,100,190]. Traditional drug-screening methods often have difficulty accurately replicating the complex physiological environment in vivo. While animal models are currently considered the gold standard for drug testing, these methods are not only costly and time-consuming but also present challenges related to specificity and ethical concerns, due to species differences between animals and humans. Additionally, the use of animal models does not align with the “3R” (reduce, refine, and replace) ethical principles in animal research [130,191].
Microfluidic-based MPSs enable the simulation of cells, tissues, organs, and even diseases at various scales in vitro [192]. In addition, microfluidics enables the flexible integration of multiple functions for a variety of drug-screening scenarios, such as drug metabolism, drug toxicity studies, and personalized medicine [7,58,189,193]. The increasing importance of drug interactions and the rising demand for screening space and combinatorial screening underscore the need for efficient methods. To address this need, integrating microfluidics with AI will optimize the drug-discovery process, reducing both the cost and time required for drug development [23,194]. In this section, we review the main types of microfluidic-based drug screening, including phenotypic screening, compound library screening, cellular activity screening, and combinatorial screening, while highlighting the role of AI in enhancing these processes. Table 2 [[195], [196], [197], [198], [199], [200], [201], [202], [203], [204], [205], [206], [207], [208], [209], [210], [211]] provides a summary of current advances in AI and microfluidics for high-throughput drug screening.
4.1. Phenotypic screening
Drug phenotypic screening is a strategy employed in drug R&D to assess the impact of various compounds on a wide array of biological effects within biological systems. Unlike target-oriented drug discovery, which concentrates on specific molecular targets, phenotypic screening evaluates the activity and effects of compounds at the cellular, tissue, or whole-organism level. This approach facilitates the identification of compounds with diverse biological effects, contributing to the discovery of new drug targets and pharmacological mechanisms—particularly in situations where disease mechanisms are complex or not yet fully understood [10,197]. The integration of microfluidics with high-throughput measurement technologies has shown great potential in phenotypic drug screening, and its synergy with AI has changed the research paradigm of this process, as shown in Fig. 7.
Microfluidic-based MPSs create 3D structures that mimic physiological conditions, integrating multiple tissue types, functional tissue interfaces, and complex mechanical and biochemical environments [104,118,212,213]. These cutting-edge platforms offer specialized in vitro models for phenotypic drug screening and investigating complex biological interactions [106,214]. Building on these advancements, researchers have successfully modeled various human OoCs, including the heart, intestines, liver, kidneys, blood-brain barrier, and nervous system [105,107,198,215]. In drug screening, tumor microenvironments (TMEs) related to vascularization introduce significant resistance. Angiogenic networks are vital for these models, as they enable the development of perfusable arrays of vascularized microtissues [197]. Organoids formed through co-culture—including those derived from breast tumors, lung cancer, skin, and neural tissue—are commonly used in phenotypic drug screening. These organoids exhibit characteristics that are similar to those of real in vivo tumors and related tissues, providing strong evidence of their relevance in human phenotypic screening [[216], [217], [218], [219], [220]]. The integration of machine intelligence can address the challenges of extracting and quantifying biochemical information from these models, thereby aiding in the development of “smart” MPSs.
AI plays a crucial role in the physiopathological construction of MPSs [221]. DL, driven by clinical data, is becoming increasingly important for assessing immune cell infiltration and cytotoxicity. Within this research framework, an automated high-throughput microfluidic platform was developed to track immune cell infiltration and cytotoxicity dynamics in 3D tumor cultures with adjustable matrix compositions (Fig. 8(a)) [199]. Using this technique, researchers identified an epigenetic drug: the lysine-specific histone demethylase 1 inhibitor (LSD1i). Deep chemical expression (DeepCE) was developed to predict drug candidates for the treatment of coronavirus disease 2019 (COVID-19) [222]. To address issues of sparsity, unreliability, and low throughput in chemically induced gene expression profiling data, the researchers introduced a novel data-enhancement method that enabled the automated extraction of valuable insights from unreliable experiments in the L1000 dataset. This phenotype-based compound-screening framework, utilizing noisy histological data, highlights the potential of AI-enhanced microfluidic drug screening for efficiently identifying promising drug candidates.
In addition to analyzing clinical data, DL has been applied to address the challenge of morphological image classification. A morphology-based CNN model was developed to identify senescent cells and characterize the pathogenesis of senescence-related diseases. This system, designated the DL senescence scoring system (Deep-SeSMo), was employed to screen a library of kinase inhibitors for drugs that can control cellular senescence, as illustrated in Fig. 8(b) [223]. Phenotypic drug screening uses high-throughput imaging techniques integrated with microfluidic systems to capture images of cells treated with drug compounds, with the goal of identifying promising candidates for treating various diseases. DL can analyze complex image features from high-content drug-treated cell images, enabling rapid and automatic classification of cell phenotypes and accurate identification of drug mechanisms of action. Moreover, AI methods for predicting molecular interactions and designing compounds with desirable properties by analyzing cellular imaging data and sequencing information are expected to enhance the drug-discovery process and improve its success rate [224].
Prior to the widespread development of MPSs, drug screening using microfluidic-based model organisms was a highly specialized approach often referred to as “animals on a chip” [189]. Through image analysis and feature-extraction algorithms, ML can identify significant features from the phenotypic data of such model organisms, thereby uncovering the biological effects of drugs. The nematode Caenorhabditis elegans and the zebrafish are commonly used as model organisms due to their close genetic similarity to humans, rapid growth and reproduction, and small size [225,226]. In one study, high-throughput microfluidics was employed to generate comprehensive brain activity maps (BAMs) of zebrafish larvae treated with a selected pharmacological agent [196]. The platform was able to extract detailed information from the BAMs to evaluate the mechanisms of action and potential therapeutic applications of compounds and to train ML models for predictive analyses. Thus, microfluidic-based model organisms serve as a framework for validating disease therapeutics, revealing clusters of drugs with intrinsic consistency linked to known therapeutic classes, and increasing our understanding of systemic neuropharmacology.
To increase the efficiency of drug screening using microfluidic platforms that rely on hydrodynamics and internally interconnected arrays, it is crucial to be able to decouple model organisms for appropriate immobilization, localization, and evaluation. Tang et al. [227] introduced an innovative fish capsule system that combines automated zebrafish capsule technology with a microdroplet microarray strategy, enabling the in vivo functional screening of both single and multi-drug therapeutics, as shown in Fig. 8(c). The platform offers a rapid and automated method for locating and immobilizing zebrafish in agarose. ML algorithms created on this high-throughput platform allow for the effective screening and analysis of small-compound libraries. More specifically, a deep-learning-assisted zebrafish phenotype screening method was developed for cardiomyopathy and cardioprotective adjunctive therapy, utilizing a dataset of 2125 labeled ventricular images for comprehensive cardiac function analysis [228]. Following a high-content phenotypic screen, cyanidin chloride was pinpointed as a potent disseminated intravascular coagulation (DIC) inhibitor. This advancement highlights the potential of deep-learning-assisted phenotypic screening in discovering promising lead compounds.
ML-enhanced droplet microfluidics also provides tremendous advantages for drug screening. By utilizing small volumes and discrete droplet containers, droplet microfluidic-based screening offers high throughput, high sensitivity, and rapid reactions with low cross-contamination. These droplet microarrays occupy a compact space while accommodating over 104-107 microdroplets in a compact area, enabling high-throughput drug reaction screening and large-scale parallel experiments [[229], [230], [231]]. An automated microfluidic platform was developed to explore droplet composition mixtures of small molecules and was validated through interactions between antibiotics and Escherichia coli ATCC 25922. This technique demonstrated excellent performance in determining the minimum inhibitory concentrations and interactions of ampicillin, tetracycline, and chloramphenicol against Escherichia coli [232]. To screen enzyme variants with extensive mutation libraries, researchers have developed an ultra-high-throughput dual-channel droplet microfluidic screening system that is capable of screening up to 107 enzyme variants per day. This system allows for the engineering of esterase enantioselectivity, which enabled the preferential production of the desired enantiomers of profens, important anti-inflammatory drugs [233]. Furthermore, the droplet screening platform can be integrated with a concentration-gradient generation module and a 2D reaction condition screening platform, which is expected to broaden the scope of droplet-based drug-screening approaches [234].
Droplet microfluidic platforms also permit the functional characterization of drug-related cytotoxicity in single and multiple cells. This approach is typically performed by wrapping the cells to be tested in droplets in order to assess the phenotypic heterogeneity of their drug sensitivity and cellular response [235,236]. Thanks to its high-throughput capabilities, droplet microfluidics produces vast amounts of valuable drug-phenotype datasets. This process can be further accelerated by using DL models to enhance the large-scale data processing involved in drug screening. Chan et al. [195] developed a microdroplet platform that uses lasers to monitor the amyloid production process, as shown in Fig. 8(d). The researchers employed a far-field camera to capture tiny spectral shifts and trained a 3D multimodal DL model on this data. Their model achieved a classification accuracy of over 95% across training, validation, and test sets, demonstrating the significant potential of DL algorithms to advance microdroplet-based drug screening.
4.2. Compound library screening
The microfluidic-based drug screening of compound libraries is an efficient, high-throughput method utilizing microfluidic technology [11,237]. It integrates the principles of microfluidics and chemical biology to quickly and accurately evaluate the effects of numerous compounds on specific biological targets or cellular functions. Compound libraries include natural products, chemically synthesized small molecules, and other sources. By screening a diverse range of compounds, candidates with significant effects on specific biological activities or disease models can be identified.
Creating drug compound libraries is a crucial step in discovering new candidate molecules for the pharmaceutical industry. A diversity-oriented synthesis strategy is employed to generate these libraries for identifying biologically active compounds. This process follows an iterative cycle of compound design, manufacturing, testing, and analysis (DMTA). Integrating a stopped-flow reactor with a high-throughput continuous platform enables the synthesis of combinatorial libraries. This approach significantly improves the efficiency of synthesizing smaller libraries by optimizing cross-reactions. Combining experimental automation with ML can further refine analytical processes, increasing the predictive capabilities, facilitating the extraction of key reaction features, and ultimately accelerating the completion of library synthesis, as illustrated in Fig. 9(a) [200]. Reutlinger et al. [238] used continuous microfluidics to generate a library of combinatorial imidazopyridine compounds. They developed an AI model based on an ant colony algorithm to predict multi-target activity. By integrating microfluidic-assisted synthesis with computer-aided target prediction, they were able to identify highly ligand-efficient antagonists for the adenosine A1/2B and adrenergic α1A/B receptors in just minutes.
Further advancements in microfluidic platforms have been made in order to study highly complex targets using synthetic or combinatorial compound libraries. For instance, droplet microfluidics has been utilized to screen the activity of DNA-encoded libraries (Fig. 9(b)) [239] and to evaluate the function of DNA-encoded compound beads [240]. The DNA-encoded synthesis sparked significant interest in combinatorial compound libraries for drug discovery, as well as automated and quantitative library screening techniques. Subsequently, a microfluidic chip was developed to facilitate multifunctional drug screening against the collective migration of cancer cells in a confined environment. The device screened a novel mechanoreceptor compound library (comprising 166 compounds) targeting restricted migration in various cancer cell lines. This screening led to the identification of three compounds that effectively inhibited pan-cancer restricted migration [241].
The combination of ML and microfluidics enables the rapid and efficient screening of extensive compound libraries, significantly improving the identification of potential drug candidates. Scott et al. [201] developed a mitochondrial parkin recruitment assay based on high-content imaging. The researchers applied multiparameter PCA and an unbiased, parameter-agnostic ML approach to analyze small interfering RNA (siRNA) screening data. In addition, they used a microfluidic platform to evaluate potentially effective drugs from the compound library at single-cell resolution, as illustrated in Fig. 9(c). In another study, researchers pre-selected 434 compounds for Sirtuin-1 inhibition from a library of 2.6 million compounds in order to showcase a virtual screening strategy that reduces large-scale screening to smaller samples. In vitro validation led to the identification of nine novel chemical inhibitors. This study illustrates how AI-driven preselection of extensive screening libraries can dramatically decrease the number of small molecules tested in vitro [202]. This framework, which combines unsupervised ML with experimental validation, also enabled the screening of new mitochondrial autophagy-inducing compounds. The researchers identified 18 small molecules from a natural compound library, discovering that these mitochondrial autophagy inducers enhanced the survival and function of glutamatergic and cholinergic neurons. This computational-experimental screening and validation workflow contributed to the discovery of effective mitochondrial autophagy modulators [203].
Integrating descriptors related to drug compounds into ML models is an effective strategy to improve prediction accuracy and efficiency. Wang et al. [204] designed a combinatorial library of over 100 000 candidate sequences and utilized a spatially constrained NN to screen antimicrobial drugs. The researchers identified three lead compounds that exhibited strong antimicrobial activity against methicillin-resistant Staphylococcus aureus. ML models that integrate vast amounts of data from combinatorial libraries and combined features demonstrate strong generalization capabilities, providing valuable examples for drug discovery, with nearly 30-times greater efficiency over traditional manual screening methods.
4.3. Cell-based assays
There is often significant heterogeneity in the response of the same type of cell to a particular drug. Conventional drug-screening methods have difficult capturing this variability, as they measure the overall response of a cell cluster. Single-cell analysis has thus gained prominence in pharmacological screening due to its potential for new discoveries, making it essential to evaluate target drugs at the single-cell level. Integrating microfluidic-based screening methods with real-time monitoring techniques enables the measurement of compound effects on cellular activity in single-cell models. For example, on-chip biochemical sensors can assess parameters such as cell proliferation, apoptosis, and metabolic activity, which can then be used to evaluate drug properties such as biological activity and toxicity [242,243].
Microfluidic platforms simplify the creation and precise control of concentration gradients for drug candidates, owing to their powerful spatial and temporal control capabilities. These microfluidic-based concentration gradient arrays provide a robust technical approach for studying cellular responses to biological, chemical, and physical stimuli in the external microenvironment, and they have been widely used in drug-screening applications such as apoptosis [244], drug-cell interactions [245], and antibiotic resistance [246]. In an earlier study, researchers presented a parallelized concentration gradient array microfluidic chip for the multi-parameter, high-throughput screening of human liver cancer cell responses. By integrating a drug gradient generator with parallel cell-culture chambers, this platform enabled simultaneous liquid dilution and diffusion, microscale cell culture, cell stimulation, and cell-labeling processes [247]. It also enabled the measurement of multiple parameters during drug-induced apoptosis in hepatocellular carcinoma cells, facilitating the rapid spectral analysis of apoptotic effectors and providing access to a wealth of information. This framework is versatile and offers valuable insights into bronchial inflammatory responses, cytotoxicity, transient receptor potential drug target screening, and the mechanisms by which contact-dependent intercellular communication influences drug efficacy [[248], [249], [250]].
Microfluidic platforms not only increase our understanding of cellular responses to external biochemical environments but also reveal the heterogeneity of single-cell signaling dynamics and mechanisms in the presence of drugs. Using machine vision and ML algorithms, biomarkers of live cell phenotypes can be analyzed and predicted at single-cell resolution, enabling the stratification of disease risk [205]. These platforms offer great potential for monitoring single-cell dynamics and gaining deeper insights into specific cell behaviors. In a study by Brouzes et al. [251], droplet microfluidics was employed to encapsulate drug reagents and single cells for droplet toxicity screening, as illustrated in Fig. 10(a). The droplet screening process on this platform is modular: First, droplets containing specific drug concentrations are optically labeled with unique codes; next, single cells are captured within the droplets and incubated for 24 h; finally, the emulsion is injected into the microfluidic chip, where fluorescence data and coding readouts are measured for each droplet. These microfluidic platforms are highly scalable, making them ideal for other applications requiring precise cell encapsulation and high-throughput droplet screening, as well as suitable for a range of omics approaches.
In a recent study, Mistretta et al. [252] developed a multi-conditional microfluidic strategy for screening drug compounds in time-resolved single cells. This microfluidic platform enabled the long-term imaging and live-cell screening of individual Mycobacterium cells. After testing fewer than 100 compounds, they identified four primary targets that met the criteria for pheno-tuning compounds (PTCs), as shown in Fig. 10(b). Another advantage of microfluidics is its ability to integrate with external fields and utilize multiphysical field-coupling effects, which significantly increases the platform’s flexibility. Zhao et al. [253] proposed a non-invasive single-cell drug-screening method based on acoustic flow, in which the acoustic pressure and shear induced by the flow improved the permeability of single-cell membranes, thereby enhancing the drug uptake process, as illustrated in Fig. 10(c). Using this dual acoustic-fluid coupling effect, propidium iodide fluorescence was successfully and efficiently delivered into cells. Loading a single myeloid leukemia mononuclear cell (THP-1) on this platform enabled the simultaneous detection of cellular responses to multiple concentrations of cytarabine chemotherapeutic agents within 30 min. This method also meets the demands of multi-drug screening or efficient combination screening in personalized therapy, greatly improving drug-screening efficiency while reducing costs.
The large-scale throughput of single-cell data generated by a microfluidic platform presents significant challenges for analyzing how different cell types respond to various external stimuli. However, the combination of high-throughput imaging and DL based on microfluidic platforms has been recognized for its considerable potential in addressing these challenges [8,224]. A microfluidic-based platform for single-cell analysis was proposed by Yellen et al. [30] to enable the flow dynamics of cell traps. By developing ML tools and integrating them with computer vision techniques, prediction of the image-based properties of over 105 clones was achieved. This advanced predictive capability allowed for the quantification of the responses of tens of thousands of single-cell-derived acute myeloid leukemia clones to drug-targeted therapies, resulting in the identification of rare drug-resistant and morphological phenotypes. Moreover, this approach can be extended to higher levels of cellular structures, such as cell pairs, organoids, and microarray live-cell fluorescence analysis, as shown in Fig. 10(d).
ML algorithms also play a crucial role in analyzing and predicting multicellular interactions. In this framework, a microfluidic-based multicellular co-culture array (MCA) provides the physical basis, and various interaction mechanisms can be simulated through a series of single-cell analyses. The high-throughput data obtained from the microfluidic platform is then used to train ML models, enabling the global prediction of parameters of interest. For example, communication between hepatocytes, antigen-presenting cells, keratinocytes, and skin fibroblasts was simulated in a platform designed to predict adverse skin reactions to drugs [206]. SVMs and PCA were used to quickly analyze the MCA data, allowing for the classification and visualization of skin-sensitizing and non-skin-sensitizing drugs. The predictive performance of this method achieved 87.5% accuracy, 75% specificity (true-negative drug-prediction rate), and 100% sensitivity (true-positive drug-prediction rate).
4.4. Drug combination screening
Combination drug therapy is widely recognized as an effective strategy for improving therapeutic efficacy by harnessing the synergistic actions of drugs with different mechanisms of action. To this end, screening a library of potentially effective drug combinations is an essential means of identifying the desired therapeutic response. Such screening can help identify drug combinations with synergistic effects, thereby improving therapeutic outcomes and reducing the risk of drug resistance.
Microfluidic systems offer an efficient and convenient platform for drug combination screening by enabling the mixing and combination of multi-drug solutions through flexible microchannel networks or droplet array designs [[254], [255], [256]]. Kulesa et al. [13] reported a high-throughput drug-screening system for nanoscale droplet emulsions that utilized parallel droplet processing for the large-scale, automated construction of chemical combinations. The system successfully predicted synergistic interactions between over 4000 drugs and 10 antibiotics, as illustrated in Fig. 11(a). This droplet-based microfluidic combinatorial drug-screening platform is highly reconfigurable, making it adaptable to complex combinatorial drug-screening processes [257]. For example, by sequentially manipulating the droplet technology, this platform was used to perform cell-based combinatorial drug screening on a 2D droplet array chip [236]. The semi-open-format screening platform enables complex, multistep drug combination screening operations—including cell culture, drug dosage, and cell viability assays—to be conducted in parallel through multiple droplet manipulations. With on-demand sample preparation, optical barcoding, and ML analysis, droplet swarms offer reliable multiplexing capabilities for microbiological applications such as antibiotic susceptibility assays [258].
By encapsulating target cells within droplets, the rapid functional screening of personalized drug combinations has also been enabled. Utilizing a Braille valve to alter droplet composition as needed, this approach allowed the collection of more than 1200 data points in a single experiment, as shown in Fig. 11(b) [259]. Microfluidic drug screening platforms with integrated Braille valves can also be combined with fluorescence-detection modules to sort droplets based on fluorescence signals following on-demand droplet generation (Fig. 11(c)) [260].
Microfluidic-based physiological and pathological models offer advanced platforms for combinatorial drug screening. These models facilitate the testing and simulation of human-level drug combinatorial responses, achieving high consistency with actual human physiology. By developing precise and robust in vitro tumor models that connect phenotype-based sphere characterization to drug activity, rapid assays for multi-drug synergistic therapeutic efficacy can be delivered to patients, as shown in Fig. 11(d) [199,261,262]. The biocompatibility, optical transparency, and physiopathological similarity of these platforms enable the analysis of multi-drug combination therapy and its mapping to solid tumor responses. However, screening all possible drug combinations through microfluidic experiments alone is not feasible due to the sheer size of the search space. Consequently, ML algorithms have been employed to efficiently and accurately identify potential synergistic anticancer drug combinations.
In combination drug screening, ML algorithms analyze the large datasets generated by microfluidic platforms to identify the effects and interactions of drug combinations [207]. Well-trained models can reduce the number of experiments required and rapidly predict potentially effective drug combinations. This capability is particularly beneficial for personalized medicine, as these models can tailor drug combinations based on patient characteristics or specific disease backgrounds. A droplet microfluidic system proposed by Davies et al. [263] was capable of generating more than 50 million combinations within a 192-well format for collaborative screening. This system improved screening efficiency through dilution, ML, and enhanced user-instrument interaction.
Interpretable ML models are essential in drug combination screening. Such models aid in selecting optimal drug combination strategies by explaining the molecular mechanisms of anticancer drug synergism. By combining accurate models with interpretable insights, interpretable ML has the potential to accelerate data-driven cancer pharmacology. Thus far, interpretable ML models have been applied to the screening of anticancer drug combinations in acute myeloid leukemia. In a dataset encompassing 133 combinations of 46 anticancer drugs in ex vivo tumor samples from 285 acute myeloid leukemia patients, the hematopoietic differentiation characteristics of therapeutically synergistic drug combinations were identified, as shown in Fig. 11(e) [209]. Although the integration of interpretable ML models with microfluidic-based OoC systems has not yet been realized, it is a promising direction for future research.
These algorithms have demonstrated high predictive performance across a wide range of tissue types and drug combinations. For example, the DL model DeepSynergy, developed by Preuer et al. [207], predicts synergistic drug combinations in cancer. This model utilizes chemical and genomic information as inputs, applies a normalization strategy to account for input data heterogeneity, and incorporates a conical layer to simulate drug synergy. DeepSynergy outperformed other ML methods by 7.2% in predicting novel drug combinations across the explored space of drugs and cell lines. The model also achieved high predictive performance, with an area under the curve (AUC) of 0.9. In another study, the ML framework comboFM was employed to predict the response of tumor-patient-derived cells to drug dose combinations in preclinical studies. This algorithm modeled environment-specific drug interactions using a higher-order tensor. The comboFM model accurately predicted drug combinations not observed in the training space, offering valuable guidance for repositioning drugs into new combinations [208].
The development of effective ML algorithms is anticipated to open new avenues for accurately and robustly predicting drug combination responses and for expanding drug combinations in individualized oncology treatment. However, experimental drug screening using a multi-dose matrix analysis is costly, particularly when identifying synergistic drug combinations in cell lines or patient-derived cells. To tackle this issue, Ianevski et al. [210] proposed an efficient ML method called drug combination response prediction (DECREASE), which enables cost-effective combinatorial screening by utilizing a minimal set of experiments to guide and accelerate high-throughput, unbiased drug combination effect screening. The model requires only a limited set of pairwise dose-response measurements for the accurate prediction of synergistic and antagonistic drug combinations. The efficiency of this method was demonstrated through the analysis of 23 595 paired drug combinations across 53 cancer cell lines, as well as malaria and Ebola infection models.
Focusing solely on synergistic effects may result in false-positive conclusions. Thus, sensitivity analyses of drug combinations are also required for ML approaches. Creating drug combination sensitivity scores (CSSs) through crossover designs and based on differences between combination and single-drug dose-response curves has been proposed as an efficient method to address this issue [211]. CSSs can be predicted using ML models and existing public datasets of known drug targets. The experimental computational method can serve as an effective drug combination screening pipeline, allowing for the systematic and economical evaluation of drug combinations. In conclusion, the synergy between ML predictions and microfluidic devices presents significant potential for reducing the cost and time constraints associated with high-throughput drug combination screening. This combination also supports broader applications in various domains of drug combination screening.
5. Application of AI in microfluidic-based drug evaluation
Drug evaluation is a crucial step in the preclinical drug-development process, aimed to ensure the effectiveness, safety, and quality of the selected drugs. Over 80% of experimental drugs continue to fail in clinical trials. Approximately 60% of these failure cases are attributed to a lack of efficacy, while another 20% are due to substandard safety [15,264]. Microfluidics enables the precise control of flow, drug concentration, biochemical gradients, and other essential substances in a scalable and high-throughput manner. As such, they can serve as valuable tools for the preclinical evaluation of therapeutic drugs. The combination of OoC co-culture systems with physiologically based PK/PD models enables the dose, toxicity, and efficacy of drug candidates to be estimated. AI can revolutionize drug evaluation by improving pharmacokinetic and toxicity predictions for more accurate assessments [265]. Table 3 [[266], [267], [268], [269], [270], [271], [272], [273], [274], [275], [276], [277], [278], [279], [280], [281], [282], [283], [284], [285], [286]] provides a summary of current advances in AI and microfluidics for drug evaluation.
5.1. PK/PD modeling and prediction
Predicting PK properties is crucial for guiding lead discovery and optimization in the early stages of drug development. The goal of PK is to study how drugs are processed within the body, predicting the PK parameters related to drug absorption, distribution, metabolism, and excretion (ADME) [287]. PD examines how a drug interacts with biological targets in the body and the mechanisms behind these interactions. It focuses on the drug’s physiological or biochemical effects, including its intensity, duration, and selectivity. Understanding dose-response relationships and the direct and indirect effects of drugs on targets is essential for evaluating therapeutic efficacy.
The scope of preclinical drug PK/PD assessment is very broad. PK parameters are mainly generated around the ADME process. In the absorption process, the maximum blood concentration (Cmax), the time (Tmax) required to reach Cmax, the blood concentration-time curve integral used to measure the overall exposure of the drug (AUC), and the bioavailability are commonly evaluated. In distribution evaluation, the drug volume of distribution (Vd) and clearance (CL) are important parameters. When evaluating the drug metabolism process, the drug half-life (T1/2) and hepatic metabolism rate are commonly used metrics that respectively indicate the time required to halve the drug concentration and the rate at which the drug is metabolized in the liver. The proportion of the drug eliminated from the body by routes such as urine or bile (i.e., excretion rate, E) and the rate of excretion through the kidneys (i.e., renal clearance, CLren) are then mainly used to evaluate the excretion process. In PD assessment, drug effects can be evaluated in terms of the maximum effect (Emax) and the half effective concentration (EC50) which respectively indicate the maximum physiological or clinical effect produced by the drug and the concentration of the drug that produces 50% Emax. In the dose-effect relationship, ED50 is used to denote the half effective dose (i.e., the dose of drug required to produce 50% Emax). T1/2(Effect) describes the time it takes for the effect to be reduced by half. In safety evaluations, TD50 and LD50 have been used to represent half the toxic dose and half the fatal dose, respectively [287,288].
Microfluidics plays a crucial role in PK/PD assessment via MPSs. These systems mimic the PK profiles of compounds by simulating the actual physiological drug exposure process [287,289,290]. Microfluidic-based organ MPSs replicate organ specificity and offer physiologically relevant dynamic properties compared with traditional static cultures and animal models [15,291].
The gut is the first organ affected by an orally administered drug, making its PK characterization essential in drug development. The systemic availability of a drug depends on its intestinal absorption and first-pass metabolism. Subsequently, the drug is metabolized in organs such as the liver and kidney, and the remainder enters the systemic circulation to exert its targeted therapeutic effect. As a consequence, intestinal microarrays, liver microarrays, kidney microarrays, and in vitro models of tissue barriers have been widely developed to assess the drug ADME process [[291], [292], [293]]. In 2010, Sung et al. [294] developed a microfluidic chip with integrated PK/PD modeling. The researchers also constructed microscale cell-culture analogs (μCCAs) of liver, tumor, and bone marrow on a chip, as shown in Fig. 12(a). This combination of PK/PD modeling with microfluidics marked an initial effort to address drug failure and species specificity issues. The reimplementation of physiologically relevant organ-organ connectivity in MPSs is crucial for quantitative physiologically based pharmacokinetics (PBPK) prediction [288,295]. Herland et al. [296] created a first-pass model of drug absorption, metabolism, and excretion in the human body. This was accomplished by fluid-coupling OoC models of the intestines, liver, and kidneys through porous extracellular matrix-coated membranes and parallel channels segregated by human-organ-specific parenchymal cellular linings. This interconnected, fluid-coupled multi-organ microfluidic platform was utilized to quantitatively predict in vitro PK parameters, and the experimental results closely matched clinical results in humans, as shown in Fig. 12(b).
Replicating the physiological environment of organ interactions in vitro is valuable for assessing PK, as it enables convenient standardization and engineering. However, in the human body, organs are interconnected through closed-loop vascular circulation. Establishing vascularized organ interconnections on a microfluidic platform provides a closer approximation of real-world conditions [287,297]. In addition, multi-organ interconnected microarrays have another advantage: They can capture the metabolism and PD of a drug as it undergoes sequential modifications in different organs. For example, Chang et al. [298] linked the liver and kidney to assess the causal relationship between the bioactivation and translocation of aristolochic acid (AA-I) and the resulting nephrotoxicity. In another study, Skardal et al. [299] reported an interconnected system of liver and cardiac spheroids, in which the effects of hepatic transformation of the parent drug on other downstream organs were assessed in detail, as shown in Fig. 12(c). Further development of multi-organ interconnected systems as a long-term goal could lead to the creation of a “human-on-a-chip” (i.e., copying organs onto a chip to mimic human function). Such a system has the potential to replace current preclinical animal experiments and provide a transformative technological pathway for drug development.
Traditional PK/PD modeling presents challenges such as high model complexity, incomplete data, and poor model generalization [266,267]. The integration of microfluidic systems and ML into PK/PD modeling is expected to improve predictive performance. This improvement is achieved through highly automated and flexible processes that leverage high-precision data acquisition, pattern recognition, and individualized modeling capabilities [300]. At present, ML in PK/PD modeling focuses on two primary tasks. First, data-driven AI algorithms are used to predict the ADME behavior of drugs, which helps accelerate drug development [301]. Second, ML is employed to optimize and dynamically adjust PK/PD model parameters [31,302].
In this section, we outline a typical process for using AI to predict the PK and PD of compounds. First, various AI algorithms are employed to predict the physicochemical properties and absorption, distribution, metabolism, excretion, and toxicity (ADMET) profiles of the compounds. Afterward, PK outcomes are forecasted using a physiologically based PBPK model. Finally, PD models estimate how changes in drug concentration affect physiological phenomena relevant to the study. Wu et al. [303] integrated graph neural networks (GNNs) with PBPK models to predict how variations in drug concentration influence gastric acid secretion and gastric pH levels. During the training of their ML model, most parameter datasets contained more than 461 data points, while the dataset for the apparent elimination rate (CLapp) contained only 98 data points. Ultimately, of the five compounds, KFP_H008 was found to have the best therapeutic effect and the longest pH > 4 retention time.
The first step in predicting human PK and PD responses when identifying drug candidates is to predict PK parameters. ML algorithms such as RF models, DTs, Gaussian processes, and DL are commonly used. Recently, these algorithms have been expanded with techniques such as transfer learning and adversarial generative networks [268,304,305]. These approaches combine ML-driven ADME properties such as permeability, ionization constant (pKa), lipophilicity, and intrinsic metabolism with a PBPK modeling framework. Chen et al. [268] developed a DT-based approach that incorporates PBPK/PD modeling and sensitivity analysis. The aim was to elucidate the optimal PK, potency, and ADME combinations for target-specific pharmacology, as shown in Fig. 12(d). Two sets of DTs were created: One uses PK parameters to determine potency, including the free peak, trough, and mean concentrations at steady state (CmaxU,ss, CminU,ss, CssU), time above half inhibitory concentration (IC50) (t > IC50), and the percentage of AUC under the initial or terminal phase; the other tree uses efficacy and ADME properties as inputs to determine the optimal molecular space for efficacy. This hybrid model provides an inverse framework, determining which PK/ADME properties are needed to achieve a specific PD response at a reasonable dose. An alternative inverse framework is provided by integrating AI-based quantitative structure-activity relationship (QSAR) models with AI-based PBPK models. This combination allows the prediction of biodistribution and PK parameters based on the physicochemical properties of the drug or the physiological information of the target organ/tissue. For instance, AI-assisted PBPK models predict key kinetic parameters related to tumors for nanoparticles by using only the physicochemical properties of the nanoparticles as inputs, as illustrated in Fig. 12(e) [269]. Fig. 13 illustrates the flow of AI in performing drug ADMET predictions and optimizing PK/PD model parameters.
In particular, the use of DL methods to construct PK/PD models allows for the direct learning of PK differential equations from patient clinical datasets [270]. The integration of these models with the pharmacological principles of dose, concentration, and effect makes them highly generalizable to untested dosing regimens. For example, using this model, drug concentrations and platelet kinetics in patients with human epidermal growth factor receptor 2 (HER2)-positive metastatic breast cancer treated with trastuzumab-emtansine (T-DM1) were successfully predicted. Transfer learning has also been effectively applied to predict ADME properties and physicochemical properties (e.g., permeability, metabolic intrinsic clearance, cytochrome P450 (CYP) inhibition, and lipophilicity), especially for targeted protein degradation (TPD) compounds, including colloidal and heterobifunctional groups [271]. Although traditional PK/PD methods are very effective in predicting drug dosage, a translation from preclinical modeling to real in vivo physiology is still lacking [306]. This issue highlights the challenges related to the availability, representativeness, and consistency of preclinical and clinical data. Therefore, standardized and easily accessible data is a crucial and urgent need for AI-based in vitro PK/PD modeling. MPSs, which mimic the flow, mechanical, and biochemical conditions of real tissues and organs, are promising complements to traditional PK/PD approaches. Advances in OoC engineering offer significant opportunities for addressing data-standardization challenges [307]. More robust prediction of PK parameters using ML models can further increase the credibility of these methods. Although this field is still in its early stages, the success stories demonstrate that combining microfluidics with ML for PK/PD modeling and prediction is a highly promising direction for development.
5.2. Toxicity and safety assessment
Approximately 40% of drug candidates are estimated to fail in development due to toxicity [264]. As a result, assessing potential toxicity and side-effects has become a critical step in the drug-discovery process. This assessment helps identify drug compounds that are safe for the human body and reduces the risk of expensive late-stage failures. Traditionally, the liver, kidneys, heart, brain, and vascular system are the main targets for drug action. Recently, assessments have expanded to include the intestines, lungs, blood-brain barrier, skin, and even specific disease tissues, such as tumors [308]. From this perspective, the emergence of microfluidic-based MPSs seems almost inevitable. These platforms provide realistic physiopathological environments for target cells, tissues, organs, and even the human body in vitro. They address the limitations of single-culture systems and animal models in terms of physiological relevance and extrapolation. Here, AI plays a crucial role in analyzing the extensive toxicity data generated by MPSs [26,309].
Traditional preclinical testing for drug-induced liver injury (DILI) relies on primary isolated hepatocytes, which have inherent limitations, such as cellular heterogeneity, functional instability, and a lack of tissue organization [310]. Liver-on-chip models utilize spherical culture and co-culture techniques to establish more realistic microenvironments. These methods help maintain and enhance hepatocyte function, leading to more accurate responses to drug-induced toxicity, as illustrated in Fig. 14(a) [311,312]. Another advantage of these platforms is their compatibility with imaging and staining techniques (Fig. 14(b)), which enable the characterization of specific drug-induced hepatotoxic injuries and their underlying mechanisms [313]. Additionally, liver microarrays with integrated biochemical sensors allow for the real-time monitoring of various cellular parameters, facilitating the rapid assessment of changes in cell viability and metabolic functions [314]. DL methods have been developed to increase the accuracy of DILI prediction [[315], [316], [317], [318]]. Ma et al. [273] developed a graph-based DL model to screen drug candidates with a high risk of DILI (Fig. 14(c)). The researchers enriched four toxicity datasets with additional attribute information, increasing the number of compounds used for DILI prediction to 15 675. This approach achieved a prediction accuracy of 81.4%. Nguyen-Vo et al. [318] developed a DILI prediction framework using a CNN and molecular fingerprint embedding features. The model was trained and tested on datasets containing 1597 and 322 compounds, respectively. The study reported an average accuracy of 0.89, and the AUC value improved from 0.90 to 0.96 compared with the previous model. Chen et al. [274] proposed a DILI prediction model based on GANs, called AnimalGAN (Fig. 14(d)), to evaluate untested chemicals. AnimalGAN generated 38 clinicopathological measurements by establishing associations between chemical exposures and clinicopathological outcomes, including factors such as chemical descriptors, treatment duration, and dose groups. The training datasets for these models are derived from established chemical databases, which may not provide the quantity and quality necessary for effective predictions. Therefore, the use of high-throughput and diverse datasets from liver microarrays is expected to improve the predictive accuracy and generalizability of DILI prediction models. In particular, GANs offer unique advantages by minimizing the need for in vitro experiments.
The kidney is one of the major targets of drug toxicity, and drug-induced kidney injury (DIKI) is a significant cause of clinical drug-development failure. Predicting the nephrotoxicity of drug candidates early in the drug-discovery process is crucial for successful drug development [308,319]. Promising biomarkers for assessing DIKI include interleukin (IL)-18, neutrophil gelatinase-associated lipid transport protein (NGAL), tissue inhibitor of metalloproteinases 2, and insulin-like growth factor binding protein 7 [319]. ML algorithms can enhance DIKI prediction to identify complex decision boundaries and classify toxic and non-toxic compounds more accurately [319]. Moreover, renal organoids can provide physiological environments that closely resemble the human kidney, thus reproducing biomarker expression in vivo. These developments lay the foundation for the integration of AI-enhanced high-throughput platforms for rapid DIKI assessment [292]. In one study, the researchers developed an in vitro model using human-induced pluripotent stem cells (hiPSCs) differentiated into proximal renal tubule-like cells. They utilized an RF model to assess cellular responses to 30 compounds in order to evaluate the prediction performance for proximal tubular toxicity [275]. An automated classifier was used to classify compounds as toxic or non-toxic to proximal renal tubular epithelial cells. The classifier achieved 99.8% training accuracy and 87.0% testing accuracy. In vitro kidney modeling and ML have shown unprecedented benefits in combining in vitro models with clinical tests and diagnostics for the reliable diagnosis and monitoring of kidney injury.
Microfluidics can potentially work in concert with ML in several other areas as well [320], including cardiotoxicity, neurological toxicity, intestinal toxicity, lung-based toxicity, blood-brain barrier-based toxicity, and the toxicity assessment of multi-organ interconnections [26,31,309]. For instance, cardiac microarrays constructed with hiPSC-derived cardiomyocytes are an effective approach for cardiotoxicity testing, as shown in Fig. 15(a) [321]. To improve the accuracy and reliability of cardiac assessment, numerous DL frameworks have been developed to improve toxicity assessment based on MPSs [322]. For example, DL frameworks such as short-time Fourier transform (STFT)-CNNs and synchro-squeezing transform (SST)-CNNs can efficiently analyze 2D data converted from mechanical beat signals, as shown in Fig. 15(b) [276]. These frameworks have achieved test accuracies of 98.55% and 99.00% in classifying drug types and distinguishing between cardiotoxic and non-cardiotoxic drugs, respectively.
The combination of immunohistostaining and real-time quantitative fluorescence monitoring can be employed to study apoptosis in brain-like organs exposed to substances such as nicotine [277]. The integration of imaging techniques and multi-field assays facilitates the acquisition of a large number of images and multidimensional data. Jimenez-Carretero et al. [278] proposed a DL model for in vitro cytotoxicity assessment. The model used microscopic images of fluorescently labeled nuclei for nucleus pattern recognition and was able to effectively predict the toxicity mechanisms of different drugs, nuclear stains, and cell lines. The main advantage of such DL tools is their ability to predict toxicity using only nucleus-stained images or fluorescence features [279,323].
Drug-induced neurotoxicity can arise through various mechanisms, including disruption of the blood-brain barrier, lipid-rich structures, energy demand, synaptic transmission, neuronal cellular structures, and neurobiochemistry. A range of assays are used to assess these toxicological effects, including stem cell platforms, exposure scenarios, functional endpoints, and predictive models [324]. One approach involved constructing predictive models using an in vitro platform comprised of multiple neuronal cells derived from hiPSCs. These models utilized a linear SVM that analyzed data from 240 neural structures treated with 34 toxic and 26 non-toxic chemicals. This SVM had a predictive value of 0.91 and correctly categorized 9 out of 10 chemicals [280]. Similarly, in vitro MPSs based on the intestine, lung, and blood-brain barrier can benefit from DL models to classify and predict physiopathological properties from large, complex amounts of data [281,282,325].
However, most of these systems are optimized for culturing individual cell types or organoid structures and lack the ability to characterize multi-organ interactions. In contrast, multi-organ microarrays have capabilities that traditional toxicity assessment models do not, such as evaluating ADMET processes [122,139], as illustrated in Fig. 16(a) [326]. This advantage is significant because these platforms are more physiologically relevant, enabling the simultaneous testing of multiple drug ADME processes and the rapid assessment of toxic responses across multiple organs. The most promising feature of multi-organ microarrays is their ability to replicate the physiological state of different individuals, paving the way for personalized medicine and toxicity assessments across diverse populations. Researchers have developed various types of multi-organ microarrays to study inter-organ interactions and their effects on toxicity, as illustrated in Fig. 16(b) [327]. However, the added complexity of inter-organ interactions presents more challenges than assessing drug toxicity in single organs [14,312]. For example, data from multi-organ microarrays often involves interactions between multiple organs and cell types; this makes the data large and complex, which can complicate data parsing. Advanced ML tools—particularly DL—have been developed to simulate organ interactions by constructing intricate network models and extracting valuable features. These advanced tools can also create time-series analysis models to predict toxic responses and optimize drug exposure regimens across various time points and doses [328], as shown in Fig. 16(c) [283].
One significant advantage of microfluidic-based in vitro MPSs is their ability to accurately model diseases as an alternative to animal models. This not only helps researchers avoid challenges associated with toxicity testing failures but also addresses ethical concerns related to species specificity. By precisely controlling the on-chip physical and chemical conditions, microfluidic systems can realistically recreate the tumor growth environment. This capability may help accelerate the preclinical toxicity testing of novel antitumor drug combinations. Additionally, advanced bio-3D printing technologies can bridge the gap between microfluidics, enabling the reproducible and automated production of complex living tissues [329]. Since the in vitro platforms used to construct TMEs, monitor tumorigenesis, and assess biophysical properties are often analyzed using images, the use of ML algorithms in this area is highly desirable. DL is especially valuable, as it supports the high-precision and efficient processing and analysis of various image data [199,284].
Thus far, ML methods have enabled the construction of a tumor biomarker knowledge graph (TBKG) that includes nodes in four categories: tumors, biomarkers, drugs, and adverse drug reactions. The TBKG can identify and elucidate potential adverse reactions to antitumor drugs. The adverse reactions of osimertinib listed in the knowledge graph were then clinically validated; the TBKG also predicted unreported adverse reactions observed in clinical cases [285].
6. AI-assisted microfluidic methods for drug manufacturing
Drug manufacturing is a key step in linking new drug designs with preclinical research. The precise synthesis of drug compounds with specific therapeutic effects is what turns “virtual” designs into real-world applications. Traditionally, these chemical synthesis reactions are carried out in bulk phase systems. Although such systems are currently widely used in drug synthesis and in the optimization of multistep drug compound synthesis, they have inherent drawbacks such as uncontrollability, inefficiency, high cost, time consumption, and high wastage rate, which limit their application in drug R&D. High-throughput microfluidics provides microscale reactors to replace traditional drug synthesis pathways, bringing unprecedented advantages.
The processing of the high-content drug synthesis data that accompanies high-throughput microfluidics is far beyond manual controllability; thus, it is a major constraint to the accessibility and diffusion of microfluidic platforms. To address this issue, AI-assisted microfluidics for drug synthesis is highly regarded [330]. In addition to generating specific drug molecules in a predictable way, AI models that can assist microfluidics in optimizing drug formulations and synthetic reaction planning have been developed, further expanding the range of AI applications. Here, we discuss the application of AI models to assist in microfluidic drug manufacturing processes, including micro/nanoparticle preparation, formulation design and optimization, synthesis reaction planning, and purification process, as shown in Fig. 17. An overview of how AI-assisted microfluidics can be used in the drug manufacturing process is provided in Table 4 [[331], [332], [333], [334], [335], [336], [337], [338], [339], [340], [341], [342], [343], [344], [345], [346], [347], [348], [349], [350], [351], [352]].
6.1. Micro/nanoparticle preparation
Droplet-based microfluidics is particularly well-suited for fabricating particles at both the micron and nanometer scales. These particles can be customized into various morphologies, sizes, shapes, and compositions, which can then be used as functionalized templates for pharmacological applications [69,353,354]. In microfluidic-based drug synthesis, the properties of drug particles are closely tied to several control parameters, such as hydrodynamic factors, reaction conditions, materials, and the operational settings of the microdevice. Precise control and optimization of these process conditions are crucial for achieving the desired drug performance, as these parameters are interrelated. ML can aid in predicting and optimizing the quantitative relationships between nanoparticle properties and control parameters, thereby facilitating the precise and controlled development of drugs [101,158,355].
Overall, the intersection of ML and pharmaceutical micro/nanoparticle manufacturing mainly consists of two main tasks [355]. The first of these is drug performance prediction, often referred to as “positive problem solving.” For instance, given a drug particle, the objective is to predict its drug effect of interest. The second task is input condition generation, which falls under “inverse problem solving.” In this case, given a desired drug effect, the aim is to predict the structure or composition of a drug particle that could achieve that effect. The use of AI models for the quantitative conformational relationship prediction of desirable particles prior to synthesis has the potential to change the difficulties of current micro/nanoparticle fabrication. Many ML methods have been applied to predict nanoparticle properties, such as the prediction of cytotoxic response, molecular loading and release, nanoparticle adhesion, nanoparticle size, and self-assembly of targeted drug-carrier particles [332,356].
Traditionally, medicinal micro/nanoparticles have mainly consisted of emulsions, microcapsules, and nanoparticles; recently, the concept of micro/nanorobots (MNRs) has been included [17,342,357]. Emulsions, a type of drug formulation, consist of a dispersion system of two or more immiscible liquids. They offer the significant advantages of bioavailability, stability, and release control [80,94]. As a powerful data analysis tool, ML can assist microfluidics to customize the structure and properties of micro/nanoparticles and reverse-engineer the input conditions, as shown in Fig. 18(a) [355,358,359]. Damiati et al. [335,336] synthesized indomethacin drug-carrying microparticles using 3D flow-focusing microfluidics and used an artificial neural network (ANN) as a machine intelligence tool to predict the size of poly(lactic-co-glycolic acid) (PLGA) microparticles, as shown in Fig. 18(b). PLGA concentration and PLGA and polyvinyl acetate (PVA) flow rate were chosen as input conditions, with PLGA particle size as the output. The established ANN predicted the PLGA droplet size highly accurately, with residuals randomly distributed within ±5 µm. Recent studies have shown that a GAN can be used to augment the dataset, thereby accurately controlling the size of the nanoparticles [360].
Nanoparticles with complex structures can be engineered for controlled release by adjusting the shell thickness and core number. However, doing so requires the precise selection of parameters such as size and thickness during manufacturing. By combining neuro-fuzzy logic and an ANN to optimize seven design variables, including water/oil (w/o) composition, polymer concentration, flow rate, and vibration frequency, Rodríguez-Dorado et al. [361] accurately encapsulated sunflower oil in calcium-alginate solution to obtain uniform spherical microcapsules of hydrophobic compounds. A recent study introduced a framework that integrates computational fluid dynamics, experimental design, and ML to predict the injectability of drug particles in particulate formulations. The study constructed a multi-physics field model to detect particle flow and clogging in the developed microinjection-needle system, and the obtained experimental data was used to train the ANN. The reverse optimization of designs with known target performance, aided by an ANN, allows for the tailoring of effective microfluidic platforms [333].
ML is widely used for the fabrication of microfluidic nanoparticles with a complex combinatorial space of structure-function relationships [[362], [363], [364]]. LNPs—the most successful drug delivery carrier candidates for clinical applications—were one of the first nanoparticles developed and studied [51]. Microfluidics allows for the bottom-up construction of liposomes and has been embraced by laboratories and commercial production in recent years because of its robust and easily scalable capabilities [365,366]. Tailoring the polydispersity and particle size distribution (PSD) of LNPs must take into account the complex nonlinear space of liposome-specific control parameters such as surface charge, phase transition temperature, and lipid concentration [338]. Rebollo et al. [334] employed an ANN model to predict the particle size and polydispersity index (PDI) of three formulations. The lipid molar ratio and process parameters—that is, the total flow rate and flow-rate ratio—were found to be the main determinants of a target size less than 100 nm and a PDI less than 0.2. The ANN model showed satisfactory agreement with the experimental values in the training phase; after model generalization, the ANN model still displayed high prediction accuracy, as shown in Fig. 18(c). In another study, Smeraldo et al. [337] developed an ANN to simulate the effect of flow rate and polymer concentration on the size of biopolymer nanoparticles, achieving an impressive accuracy of 98.9%, as shown in Fig. 18(d).
Microfluidics provides the opportunity to synthesize drug-loaded LNPs with uniform particle sizes. The extensive dataset of characterization parameters generated from this process provides an excellent foundation for developing ML models. Wu et al. [341] employed rapid nanoprecipitation for the continuous production of chitosan nanoparticles (CNPs) and used 12 ML algorithms to investigate the hidden relationship between the CNPs and reaction-independent variables (i.e., particle size and dispersibility index). Among these, RF, DT, and bagging algorithms were found to have a prediction accuracy of 90%. Wang et al. [367] demonstrated the potential of AI for the rapid synthesis of anticipated drug nanoparticles by using a high-throughput microfluidic platform and a particle-size-dependent ML model to predict candidate synthesis schemes for LNPs of a desired particle size. A high-throughput experimental platform was constructed by integrating automated modules, which provided 672 experimental protocols for the synthesis and characterization of drug-loaded LNPs in 40 h. The prediction performance of four ML models—namely an RF, a back propagation neural network (BP NN), an SVM, and Gaussian process regression (GPR)—was compared.
Similarly, a wide range of ML frameworks are available for predicting the properties of drug nanocrystals, as illustrated in Fig. 19(a) [339,340,368]. The nanocrystallization of active pharmaceutical ingredients (APIs) includes various drug solid forms such as polycrystalline materials, eutectic/salts, and solvates. Prior to clinical trials, it is necessary to fully understand the drug candidate information, including potential crystalline forms, physicochemical properties, and manufacturing approaches. This knowledge is vital for implementing monitoring, modeling, and control strategies based on high-throughput microfluidic big data. ML algorithms have shown considerable potential in predicting nanocrystal physicochemical properties, crystal structures, and co-crystal conformation [369,370].
The ability of ML models to predict the solubility, melting point, lipophilicity, flowability, and filterability in crystallization processes is extremely valuable for drug development, particularly regarding pharmacokinetic properties, drug-target binding rates, and product quality control, as illustrated in Fig. 19(b) [339,371]. Solubility prediction in ethanol using a combination of ANN and random decision forest (RDF) algorithms for APIs has been successfully achieved with high accuracy [372]. Boobier et al. [373] accurately predicted the physicochemical relationship between solubility and molecular properties across different solvents by combining ML with computational chemistry, achieving a log accuracy of ±0.7. DL algorithms such as ANNs, GNNs, graph convolutional networks (GCNs), and DNNs can effectively capture nanocrystal structural information and predict physicochemical properties, including the solubility, tendency to crystallize, and powder flowability of drug solute systems, as illustrated in Fig. 19(c) [[374], [375], [376], [377]].
AI models can also be applied to predict the properties of polycrystalline materials and eutectic systems. The Crystallography Open Database is widely used as a training dataset for AI-assisted crystal analysis, covering various crystal materials, including inorganic crystals, biomolecules, and proteins. SVM and NN algorithms have successfully predicted stable polycrystalline forms of oxacillin and AZD1305, as illustrated in Fig. 19(d) [378]. In addition, combining DL with a droplet microfluidic platform has enabled the screening of APIs under antisolvent crystallization conditions [340]. For example, indomethacin crystals with different morphologies were generated in hydrogel droplets and captured through optical imaging. DL algorithms were developed to recognize and classify large-scale images of the crystals, which were then used to guide the scale-up preparation of indomethacin crystals.
6.2. Formulation design and optimization
A key role in the drug development process is played by drug formulation, which involves the drug composition, ratio, formulations, and preparation processes. The design and development of an advanced pharmaceutical formulation is a complex process, and this complexity stems from a number of factors, including the need to consider a variety of parameters related to the drug, the inert excipient, and the manufacturing process in a high-dimensional design space [379,380]. The integration of microfluidics and ML holds potential to enable pharmacy scientists to map the relationship between the composition and performance of advanced drug formulations to enhance a priori formulation design [381].
Given the complexity of formulation development, ML is a valuable tool for modeling and analyzing formulation selection and optimization [343,359,382]. Han et al. [383] developed a DNN model based on the maximum dissimilarity algorithm with a small group filter and representative initial set (MD-FIS) selection algorithm to predict formulations of oral disintegrating tablets (ODTs) manufactured via the direct compression process. This approach was a novel application of DNNs combined with an enhanced dataset selection algorithm for formula prediction on small datasets, which were divided into a training set (105 cases), a validation set (20 cases), and a test set (20 cases). The DNN model achieved 85.60%, 85.00%, and 80.00% accuracy in the training set, validation set, and test set, respectively. In another study, three ML tools (i.e., RF regression, symbolic regression, and an ANN) were used to predict the powder formulation characteristics and performance of a continuous pharmaceutical blender using the revolutions per minute (RPM), mixing blade area size, wall friction angle, and feeding rate as the input features. The ANN was shown to be the best predictor in terms of predictive power, with R2 = 0.97. After training using a single-component mixture, the ANN predicted four sets of paracetamol powder formulations at three rotational speeds of 250, 350, and 450 RPM, with an average error of only 1.4% [384]. Overall, ML-modeling-assisted formulation development can expedite the identification of the optimal range of process parameters and facilitate the assessment of critical features in the manufacturing process.
Nanomedicine is another area of formulation development that offers extensive opportunities for ML. The use of microfluidics in the development of nanomedicines allows for precise definition of the composition, structure, size, and factors [39,385]. In addition, data acquisition tools such as high-throughput imaging, Raman spectroscopy, and infrared spectroscopy are conveniently integrated with microfluidics, providing effective tools for high-quality data collection [8,386,387]. Sheng et al. [345] developed a method for intelligent characterization of protein particles on a digital microfluidic chip combined with microscopic Raman spectroscopy, as shown in Fig. 20(a). Five ML models (DT, discriminant analysis, SVM, enhancement classifier, and NN) were trained using 220 Raman spectra of protein particles in droplets, and the well-trained models successfully classified protein particles induced under eight stress conditions. The prediction accuracy was 100% under five degradation conditions: oscillation, freezing-thawing for 10 cycles, 80 °C, pH 3, and 2000 kilolux-hour.
Other examples of the use of similar frameworks include the prediction of dissolution profiles [379], the quantification of multiple ingredients in drug formulations [388], and the prediction of in vitro tablet drug release [389]. These models accurately predicted physicochemical properties such as the particle size, targeting efficiency, slow-release properties, and overall toxicity of the desired particles, using molecular formulation parameters such as composition, content, chemical structure, and degree of polymerization as inputs on both continuous and droplet microfluidic platforms, as shown in Fig. 20(b) [344,390,391]. Inverse problems in the form of forward solutions can likewise be solved using this framework. Kimmig et al. [343] proposed a combination of multilayer graphical CNNs and fully connected NNs to accurately predict the particle size of polymer nanoparticles using a dataset of 3753 nanoparticle formulations, requiring only the polymer structure, degree of polymerization, and nanoprecipitation parameters, as shown in Fig. 20(c). Although such efforts are still in their infancy, the development of such combined platforms and applications based on data sharing will facilitate experts and practitioners in different fields and interdisciplinary collaborations to accelerate the process of microfluidic formulation development in a data-driven manner. This paradigm shift will ultimately allow for nanomedicine similar to autopilot labs [392].
6.3. Synthesis reaction planning
Synthetic reaction planning is critical for the discovery and manufacture of drug compounds, especially small-molecule drugs. This process involves designing efficient and controllable reaction pathways that minimize the number of steps and optimize reaction conditions, ultimately improving the yield and purity of target molecules. In addition, rational synthetic planning enables researchers to explore and optimize the chemical structure of drug molecules and improve their pharmacological and biological activities. Synthetic planning benefits from the superior data-prediction capabilities of advanced ML algorithms, since the design space consisting of the number of possible compounds and routes is almost impossible to traverse manually [348].
AI-assisted microfluidic chips, commonly known as smart microreactors, integrate the advantages of intelligent algorithms, high-throughput controlled flow and measurement platforms, and numerical computation tools. They simulate and reveal the physical processes of synthetic reactions, while enabling the study and optimization of synthetic reaction conditions, routes, and steps in a visual and controlled manner, as shown in Fig. 21(a) [346,349]. Inverse synthetic planning is one of the main tasks of computer-aided synthesis planning (CASP), as shown in Fig. 21(b) [393,394]. Solving inverse problems is a key capability of ML. Beginning with a target molecule, a range of potential synthetic pathways and reaction sequences can be deduced in reverse to facilitate the design of efficient synthetic routes [395]. The Chematica software was used to autonomously design synthetic pathways for eight structural targets, and all of these computer-planned routes were then successfully implemented in the lab [396]. DNNs have also been used to resolve reactive conflicts. Trained on 3.5 million reactions, a graph-based NN model achieved 95% accuracy in reverse transcription synthesis and 97% accuracy in reaction prediction on a validation set of nearly 1 million reactions (Fig. 21(c)) [397]. This approach frees chemists from manual labor by breaking down complex molecular structures to generate a series of hypothetical synthetic routes, allowing for more efficient and focused design and experimentation in drug synthesis. ML models that incorporate physical constraints demonstrate increased learning efficiency and generalization capabilities [393]. Moreover, incorporating reaction templates developed by chemists into machine algorithms can significantly improve model interpretability [398,399]. For example, Chen et al. [399] introduced a DL model by customizing automatic extraction algorithms and reaction templates, which were validated in inverse synthesis, reaction result prediction, and atom-to-atom mapping tasks (Fig. 21(d)).
Recommended reaction conditions are another critical aspect of CASP. Traditionally, the identification of optimal or acceptable reaction conditions has relied on time-consuming empirical screening performed by experienced chemists. With the aid of AI, historical experimental data can now be used to train intelligent systems capable of inferring suitable reaction conditions and suggesting a diverse set of reactions in an objective and efficient manner [400]. Forward reaction prediction is one of the fundamental tasks in CASP. It involves designing one or more synthetic pathways for a known compound structure, starting from simple or commercially available starting materials. Unlike retrosynthetic analysis, forward synthesis planning focuses on the step-by-step construction of complex target molecules from basic precursors. In addition to quickly generating a wide range of possible synthetic pathways and reaction sequences based on predefined rules and literature data, AI models can predict byproducts, estimate yields under various conditions, and explore novel areas of chemical space [351,393].
Optimizing chemical reactions is essential, as unoptimized chemistry can be costly in terms of both time and material consumption [347]. Zhou et al. [351] developed a deep reaction optimizer that combines deep RL with chemical domain knowledge to guide the optimization of reactions using a microfluidic droplet reactor. The model can optimize not only reaction yield but also parameters such as selectivity, cost, and purity. Since complex multistep chemical reactions are common, an ideal automated synthesis platform should be capable of independently planning and executing synthetic routes under conditions suitable for large-scale production. Volk et al. [350] introduced AlphaFlow, a self-driven fluidic laboratory designed for the autonomous discovery of complex multistep chemical reactions, as illustrated in Fig. 22(a). AlphaFlow integrates RL with a microdroplet reactor to execute programmable sequential reaction steps, including phase separation, washing, and other key processes. In the study, a multistep reaction pathway—comprising up to 40 parameters—was successfully identified and optimized, using the synthesis of core-shell nanoparticles as a case study. This work not only highlights the potential of RL in handling complex multistep chemical reactions but also enables the prediction of time-sensitive hidden states associated with unstable intermediates.
The ultimate goal of synthetic planning is to develop intelligent systems that autonomously close the loop, eliminating the need for manual configuration. In this new paradigm, reverse synthesis, condition selection, formulation development, experimental configuration, and process execution will all be fully automated, forming the concept of the self-driving lab [346,347,392,401]. Coley et al. [348] took a useful step toward exploring the possibility of the complete automation of drug conception to execution by combining AI-driven synthesis planning with a robot-controlled microfluidic experimental platform. Their system, which consisted of a synthesis planning module and a robotic flow chemistry platform, was validated through the synthesis of 15 drugs or drug-like substances, as shown in Fig. 22(b). The study integrated CASP, expert refined chemical recipe generation, and a robotic flow chemistry platform to form an automated, scalable, and reproducible synthetic planning route. This development strategy allows drug synthesis experts to focus on new knowledge discovery, freeing them from repetitive, monotonous tasks and assigning the task of planning synthesis routes to machine intelligence.
7. Emerging AI-driven drug delivery based on microfluidics
Drug delivery involves targeting therapeutic agents to specific sites in the body in order to increase treatment effectiveness. However, traditional systemic delivery methods present challenges in controlling the timing, dosage, and rate of drug release, which can result in fluctuating drug concentrations and diminished efficacy. These conventional approaches also suffer from issues such as poor bioavailability, difficulty in crossing biological barriers, inadequate targeting, and the need for frequent dosing, all of which can affect patient adherence. As attention shifts toward next-generation therapeutics, including proteins, peptides, and nucleic acids, it is essential for drug delivery technologies to evolve accordingly [402]. Microfluidics allows for the precise control of fluid within micron-scale channels, paving the way for innovative drug carriers such as LNPs and microscale robots. It also facilitates advanced delivery platforms, including microneedle-based systems, microfluidic implants, and wearable devices with integrated sensors for real-time monitoring. When combined with AI, microfluidics significantly improves the design and optimization of drug delivery systems, representing a major advancement in drug delivery strategies. By harnessing the strengths of microfluidics and AI, new possibilities can be opened up for the intelligent design of these systems, thereby improving delivery pathways and enabling real-time monitoring and regulation.
7.1. Intelligent drug delivery system design
Microfluidic drug delivery systems can be classified into three main application types: in vitro, in situ, and in vivo. In vitro microfluidic systems operate in controlled laboratory settings, allowing researchers to study drug behavior and simulate drug release under various physiological conditions [403]. In situ systems are designed for localized therapy, delivering drugs directly to specific sites within the body. In vivo microfluidic systems administer drugs within a living organism, enabling real-time therapeutic interventions [404]. According to the target and delivery characteristics, these systems can also be categorized into targeted drug delivery systems (TDDSs), controlled-release drug delivery systems (CRDDSs), and smart drug delivery systems (SDDSs). The goal of TDDSs is to deliver drugs precisely to specific cells or tissues. These systems achieve high selectivity by using targeting molecules—such as antibodies, ligands, or small molecules—that recognize specific biomarkers [365]. CRDDSs aim to stabilize drug release over extended periods, thereby reducing the frequency of administration. They are particularly useful in managing chronic diseases and supporting long-term therapies. SDDSs can modulate drug release in response to physiological changes, such as variations in pH, temperature, or enzyme activity. These systems enable personalized and dynamically adjustable therapeutic regimens tailored to evolving disease states [405,406].
To successfully overcome biological barriers in vivo and achieve targeted drug delivery at the cellular and tissue levels, a full understanding of the interaction mechanisms between drug delivery vehicles, extracellular matrices, target tissues, and cross-barrier transport is essential [407]. Many TDDSs have been developed for specific targets using microfluidic platforms. These systems include nanoparticle-based delivery systems, liposome-based systems, polymeric carriers, antibody-drug conjugates, exosomes, and extracellular vesicles (EVs) [51,408,409]. AI models play a key role in optimizing the production process to improve the efficiency and consistency of drug delivery systems. This is especially important because the therapeutic efficacy and delivery properties of nanodrug systems are often directly related to factors such as particle size and structure [29,364]. Furthermore, ML algorithms are applied to predict drug-carrier interactions and to refine carrier designs and drug-release mechanisms. ML models enable the rapid screening of effective drug delivery systems, helping to identify the optimal delivery solution. Additionally, the application of AI to simulate drug distribution and effects within individuals allows for the prediction of drug delivery system performance and the provision of personalized drug delivery solutions. Importantly, AI processes data from microfluidic devices or sensors in real time, enabling the adjustment of drug delivery parameters and driving innovation in novel TDDSs [332,410,411].
The design of drug delivery systems requires careful consideration of various factors to ensure drugs’ efficacy, safety, and reliability. A thorough understanding of a drug’s properties is essential in this process. Key design objectives include the physicochemical properties of the drug, PK/PD in vivo, and the drug’s release kinetics [412]. Models such as LightGBM, classification and regression trees (CARTs), GPR, and MLP are used to uncover the physicochemical relationships between factors such as solubility, stability, molecular weight, hydrophilicity, structural features, and solvents [[413], [414], [415]]. Hybrid models combining mechanistic modeling with ML have been employed to predict how drug loading affects the release of levonorgestrel from spray-dried poly(levulinic acid) particles [416]. In this case, data on the actual drug content (ranging from 6% to 52%), particle size, and polymer crystallinity (ranging from 0 to 15%) was used to construct the models. The results demonstrated that drug diffusion and polymer degradation had a more significant impact on drug release in particles with low drug content. Drug delivery systems must specify the targeted cell, tissue, or lesion region with the corresponding targeting mechanism. Subsequently, the design of drug delivery systems is based on the choice of carrier, while considering the biocompatibility and degradability of the carrier material.
LNPs have significantly advanced the use of liposomes as nanodrug carriers due to their distinctive hollow structure. Hunter et al. [417] applied this methodology to extract data-rich phenotypic fingerprints from cellular imaging, utilizing ML to identify key features correlated with enhanced delivery. This methodology has facilitated a more profound comprehension of advanced cellular and endocytosis analyses of intracellular delivery, significantly improving the efficacy of LNP-based messenger RNA (mRNA) delivery, both in vitro and in vivo, as shown in Fig. 23(a). The integration of DL with combinatorial chemistry has led to the development of the AI-Guided Ionized Lipid Engineering (AGILE) platform shown in Fig. 23(b) [418]. These studies all highlight the ability of ML to accelerate the development of customized LNP drug delivery systems [379,390].
EVs are also gaining attention in drug delivery due to their role in regulating intercellular communication and molecular transport [[419], [420], [421]]. Patient-derived EV isolates can be easily obtained from blood, urine, and saliva using microfluidic technology [408]. Applying AI to precisely define and optimize EV regulatory functions, as well as to guide the rational design of EV-based drug delivery systems, presents a significant opportunity to improve drug efficacy by controlling physiological responses in cells, tissues, and organs [422]. Moreover, generative AI enables the customization of targeting ligands and of precise drug delivery vectors associated with EVs [423]. The AI-enabled EV drug delivery approach has the potential to increase therapeutic efficacy, reduce adverse effects, and significantly advance the development of targeted and personalized medicine.
The key to SDDSs lies in the prediction and control of drug-release kinetics—more specifically, the rate and pattern of drug release in vivo—along with a clear understanding of the release mechanism (e.g., dissolution, diffusion, or degradation). This includes managing immediate, delayed, or responsive release profiles [25,411]. Techniques such as DL and regression analysis allow for the simulation of drug-release dynamics in the body and the prediction of drug-release profiles. In addition, AI models offer the advantage of optimizing the design parameters of the drug delivery system—such as carrier material, drug loading, and release mechanisms—to achieve the desired release rate and timing. Bannigan et al. [424] evaluated the performance of 11 different ML algorithms in predicting the drug-release scores of long-acting polymer injectables. These models were trained using a database of in vitro release studies, which contained 181 drug-release characteristics and 3783 data points across 43 unique drug-polymer combinations. The study found that the light gradient boosting machine (LGBM) model accurately predicted partial drug release. In a different study, spectral measurements, process data, and key material properties were used to support three ML models: an ANN, an SVM, and an extremely randomized tree (ERT)—to predict the in vitro dissolution of extended-release tablets. The impact of factors such as the API and matrix polymer PSD on the drug-release rate were also evaluated in detail [425]. ANNs have also been applied to predict the in vitro release characteristics of various extended-release tablet prescriptions, including the dissolution profiles of enteric-coated tablets and the release and control of diazepam from melt-deposited printed tablets [426], as shown in Fig. 24(a) [424].
ML models can enable the identification of complex patterns in drug release processes, release profiles, and response mechanisms under varying conditions (Fig. 24(b)) [426], leading to the development of smart drug-release systems [427]. Suryavanshi et al. [428] investigated the performance of temperature-responsive self-folding feedstocks for autonomously monitoring drug release under real physiological conditions. This was achieved using four-dimensional (4D) printing—an emerging drug delivery system manufacturing technology—in combination with ML modeling. Parameters such as shape solidity and recovery rate were used to train the ML algorithms. The temperature-responsive contraction and swelling properties were confirmed, with releasing almost 100% ± 4.19% of acetaminophen (paracetamol, PCM) in gastric pH medium over 4 h, with PCM being used as a model drug from this programmed 4D-printed construct. Another study integrated AI with the microfluidic-based synthesis and characterization of nanophotosensitizers, offering a relatively aggregated model for developing effective photodynamic therapy applications for cancer treatment [429]. These examples provide ample evidence of the ability of AI models to correlate quantitative relationships between drug delivery design parameters and release kinetics, providing a valuable tool for the development of SDDSs.
7.2. Drug-administration route
The study of drug delivery routes is a key aspect of drug development and clinical application, with the primary goal being the effective and safe delivery of drugs to the target site. Traditional drug delivery methods, such as oral, injectable, and topical administration, have shown significant results in various therapies. However, their limitations—such as issues related to intestinal and hepatic absorption, degradation, pain, and the risk of infection—have driven ongoing efforts to explore and develop novel delivery methods. To date, new drug delivery routes based on microfluidic principles—such as microneedle patches, transdermal drug-release patches, and various wearable drug delivery devices—have been developed for transdermal, implantable, and in situ drug delivery applications, as illustrated in Fig. 25(a) [59,430,431]. These advanced drug delivery systems integrate micropumps, micro-valves, micro-reservoirs, biochemical sensors, microfluidic delivery systems, and AI control algorithms, enabling a closed-loop process for intelligent drug delivery, release, monitoring, and feedback [432].
Transdermal drug delivery systems are an advanced drug delivery method in which the drug is transported through the stratum corneum of the skin [433]. This method avoids the first-pass effect of the digestive system and provides sustained drug release, leading to stable blood concentrations. Common forms of transdermal drug delivery include microneedle arrays and microfluidic drug-release patches. These systems enable efficient and customized drug delivery by precisely controlling the rate and amount of drug released [189,404]. The incorporation of customized AI algorithms further enhances the intelligence of these systems, facilitating personalized drug delivery and dynamic adjustments through data analysis, algorithmic optimization, and the real-time monitoring of drug-release parameters. The integration of AI with microfluidics not only improves the efficiency and safety of drug delivery but also advances the development of smart medical devices [411,430].
AI can be used for the fabrication of drug-delivering microneedles, which enables the optimization of both the fidelity of microneedle-array additive manufacturing and the morphology of the microneedles themselves [434]. Bagde et al. [435] achieved optimal microneedle printing fidelity and morphology by using a CNN to analyze images of dissolvable microneedle patches for microneedle fabrication, as shown in Fig. 25(b). The CNN’s input layer consisted of more than 500 high-resolution microneedle images, including 27 manually annotated images to categorize them. A polynomial relationship between the patch release and in vivo uptake demonstrated that the soluble microneedle patches exhibited sustained transdermal drug delivery and release performance, demonstrating the capability of DL in microneedle manufacturing. ML’s ability to analyze and recognize patterns in high-dimensional datasets enables it to learn and optimize the manufacturing parameters of microneedle arrays, as well as drug formulations. Algorithms such as SVM, RF, ANN, KNN, and LR have been used to predict critical process parameters, including mechanical properties, printability, extrusion temperatures, and printing temperatures of drug filaments [436].
Implantable drug delivery systems (IDDSs) enable precise drug release and long-term therapy, offering improved bioavailability and durability by delivering drugs directly to specific sites in the body. IDDSs based on microfluidics offer a new approach to designing and optimizing drug-release mechanisms, thanks to their ability to be precisely controlled at the micron scale [59,437]. These systems are advantageous for chronic disease management, oncology, and the treatment of central nervous system disorders, where sustained drug release, localized drug delivery, and penetration of biological barriers are critical. Applications of IDDSs include the stable release of insulin for diabetes management, TDDSs for tumor treatment in cancer, and direct drug delivery across the blood-brain barrier for Alzheimer’s disease [438]. Surini et al. [439] developed a controlled-release system for implantable protein drugs using a multi-ionic complex and sodium hyaluronate and evaluated the drug release from microspheres, with insulin as a model drug. An ANN model was also employed to predict the insulin-release rate constant in relation to the polymer mixing ratio, total particle weight, and compression pressure. The model successfully identified the non-linear relationship between the influencing factors and the release response. In another study, Benko et al. [440] employed an ANN to predict drug dissolution, demonstrating that drug release from IDDSs can be customized through interactions between the drug and its carrier. The model incorporated compression pressure, excipient dosage, and the API as independent variables, while the hardness, porosity, and release rate served as inputs. The release rate and release index were used as the outputs of the ANN model. The predictions indicated that API-excipient interactions have a considerable effect on drug release by retaining the drug in the matrix, demonstrating the potential of ANN-based modeling in the design of implantable systems with tailored dissolution characteristics.
MNRs, a specialized category of IDDSs, can precisely navigate, identify, and target specific cells and tissues in vivo. Their small size facilitates highly accurate drug delivery at the cellular level. AI plays a crucial role in various aspects of MNR systems, including materials, design, fabrication, propulsion, sensing, navigation, control, and collective behavior. For instance, RL and genetic algorithms have been employed to optimize propulsion strategies for various micro-swimmers, as shown in Fig. 25(c) [441]. Deep RL establishes a correlation between the induced magnetic field and the robot’s motion, allowing for the navigation of a soft, magnetic microrobot [442]. These capabilities enable MNR-based drug delivery systems to navigate the intricate pathways of the central nervous system and overcome the blood-brain barrier, thereby increasing therapeutic efficacy while minimizing off-target effects. AI can also improve the automation, adaptability, and immune-response modifiability of MNRs, facilitating the precise delivery of dosages to targeted areas. The integration of AI into IDDSs based on MNRs holds the potential to revolutionize the design and operation of next-generation SDDSs [342].
8. Challenges and perspectives
Advancements in drug R&D have been greatly accelerated by the integration of microfluidics and AI. Microfluidics provides a high-throughput platform for drug-discovery pipelines by enabling the precise manipulation of small volumes of fluid and the seamless integration of multiple modular components. Meanwhile, the advanced data analysis and learning capabilities of AI are leveraged to identify valuable patterns and insights from large experimental datasets. Consequently, AI is extensively employed across various stages of drug development, including early drug discovery, preclinical evaluation, manufacturing, and drug delivery processes. The integration of these two technologies not only shortens drug development cycles and reduces costs but also paves the way for personalized drug development and improved treatment for complex diseases. This review emphasizes that the deep convergence of AI and microfluidics will provide a transformative tool for all stages of the drug development pipeline. Nevertheless, several challenges remain before the seamless convergence of AI and microfluidics can reach full maturity and be comprehensively applied to the drug-discovery process.
8.1. Microfluidic data accessibility and availability
The first challenge in integrating AI into the drug-development pipeline is ensuring access to the high-quality, accurate, large-scale data necessary for building AI models. Data access encompasses both accessibility and availability—that is, the accuracy, consistency, and ease of obtaining data. While there are excellent publicly accessible databases, issues such as data reliability concerns and limitations on open access due to privacy and data-protection laws still pose significant obstacles. In this context, integrating microfluidics with out-of-field, modular detection components presents a promising opportunity for acquiring high-quality and high-throughput data. Researchers can easily customize microfluidic systems to take advantage of their efficient mass transfer, chemical reaction capabilities, and highly parallelized processing features, enabling the collection of labeled data of interest. For example, droplet microfluidics plays a key role in fluorescence coding and phenotypic screening during drug concentration testing [230,251]. Coupling these systems with components such as MS, Raman spectroscopy, and imaging techniques makes it possible to monitor both the spatial and the temporal distributions of drugs and flows, thereby providing multidimensional and reproducible data [8,162,345]. Furthermore, the complexity of OoCs continues to increase, as is needed to more accurately recreate in vitro models that better reflect the human body’s physiological conditions. As a result, the scale and credibility of data on drug dose-response relationships, ADMET, multi-omics, and even personalized data will be significantly enhanced, which will help meet the accessibility and availability requirements for building robust AI models.
To improve data sources for microfluidics, it is essential to develop robust strategies that increase the reproducibility and openness of microfluidic data for AI applications. The stability and reliability of data from high-throughput microfluidic platforms remain a challenge, primarily due to varying practices among stakeholders. This has resulted in a lack of standardized protocols for microfluidics, including inconsistent terminology, processes, and systems [307]. Importantly, microfluidic-based MPSs are expected to complement or even replace animal models as the gold standard for drug testing. Therefore, MPS standardization is urgently needed to ensure that drug screening and evaluation data are reproducible and translatable [443]. This would help avoid costly, time-consuming steps to produce high-confidence, standardized data.
Another challenge is that OoCs are not as efficient or cost-effective as other microfluidic methods, and the amount of data obtained from an MPS is often limited. In this context, a small-sample strategy becomes valuable for AI model development using MPSs. Techniques such as data augmentation, GANs, and transfer learning offer promising solutions to increase the learning capabilities and predictive performance of models, particularly in cases where data is scarce.
8.2. Demands and limitations of interpretable AI
Drug development is a complex and high-risk process that spans multiple stages, from identifying drug targets to conducting clinical trials. Given the complexity and stringent accuracy requirements, the interpretability of AI models is especially crucial [154]. AI models are commonly used in drug screening to predict the activity or toxicity of specific compounds. To effectively leverage these predictions, researchers must understand the rationale behind the model’s outputs. This interpretation not only helps validate the model’s reliability but also enables scientists to uncover potential biological mechanisms or identify new drug targets. In this way, interpretable AI fosters transparency and trust, making it easier for scientists to accept AI-driven recommendations and better design subsequent experiments. However, many AI models in use today, particularly DL models, are inherently “black box” in nature. For example, while DNNs may excel at predicting drug side effects, their complex architectures and large number of parameters make it difficult to interpret their decision-making processes [278]. This lack of transparency can undermine researchers’ ability to access the reliability and scientific validity of predicted results, potentially impacting drug-development decisions.
Most applications of microfluidics and AI in the drug-development pipeline focus on post-experimental data analysis, such as predicting the structure or composition of drug particles related to drug efficacy [332]. However, using AI to control microfluidic devices could significantly advance drug development by enabling next-generation, data-driven autonomous drug-testing platforms, often referred to as “self-driving labs” [392]. One major challenge in this direction is device compatibility, so integrating AI technology with existing microfluidic devices and experimental processes may require further development to address challenges related to device compatibility and data interfacing. Another challenge is the complexity of microfluidic drug development, which involves physical, chemical, and biological processes across multiple scales. AI models may struggle to fully capture the intricate details and interactions of these processes. Finally, in practical applications where real-time data processing and feedback are essential, the real-time performance and responsiveness of AI models may become a limiting factor.
9. Conclusions
In summary, the convergence of AI and microfluidics holds immense value throughout the drug-development pipeline. By integrating these technologies at three key levels—namely, data, algorithms, and hardware—autonomous intelligence can be achieved. This synergy has already proven beneficial in early drug discovery, high-throughput drug screening, drug ADMET assessment, drug manufacturing, and the design of intelligent drug delivery systems and routes. Together, these technologies are revolutionizing drug R&D, significantly improving both the efficiency and the success of drug development. As AI and computational power continue to evolve, this convergence will drive further advancements in drug discovery, allowing for the management of increasingly complex datasets and delivering more precise design and optimization solutions. Simultaneously, advances in microfluidics will increase automation and efficiency in experimental processes, enabling broader drug screening and more personalized development approaches. When combined with emerging technologies such as biochemical sensors, e-skin, wearables, and gene editing, the synergy between AI and microfluidics promises to make a profound impact on drug discovery, mechanistic studies, and clinical trials, paving the way for precision medicine.
CRediT authorship contribution statement
Du Qiao: Software, Methodology, Formal analysis, Data curation, Conceptualization, Writing - original draft, Visualization, Validation. Hongxia Li: Writing - review & editing, Supervision, Funding acquisition. Xue Zhang: Writing - original draft, Visualization, Software, Resources, Data curation. Xuhui Chen: Writing - original draft, Visualization, Software, Data curation. Jiang Zhang: Writing - original draft, Visualization, Software, Data curation. Jianan Zou: Writing - original draft, Visualization, Software, Data curation. Danyang Zhao: Supervision, Methodology, Validation. Weiping Zhu: Writing - review & editing, Validation, Supervision. Xuhong Qian: Writing - review & editing, Supervision. Honglin Li: Writing - review & editing, Supervision, Funding acquisition.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work was supported by the National Natural Science Foundation of China (82425104) and the National Key Research and Development Program of China (2022YFC3400501).
VamathevanJ, ClarkD, CzodrowskiP, DunhamI, FerranE, LeeG, et al. Applications of machine learning in drug discovery and development. Nat Rev Drug Discov2019; 18(6):463-77.
[2]
ElviraKS. Microfluidic technologies for drug discovery and development: friend or foe? Trends Pharmacol Sci 2021; 42(7):518-26.
[3]
HayM, ThomasDW, CraigheadJL, EconomidesC, RosenthalJ. Clinical development success rates for investigational drugs. Nat Biotechnol2014; 32 (1):40-51.
[4]
SeokJ, WarrenHS, CuencaAG, MindrinosMN, BakerHV, XuWH, et al. Genomic responses in mouse models poorly mimic human inflammatory diseases. Proc Natl Acad Sci USA2013; 110(9):3507-12.
[5]
DuggerSA, PlattA, GoldsteinDB.Drug development in the era of precision medicine. Nat Rev Drug Discov 2018; 17(3):183-96.
[6]
QiaoD, LiHX, ZhuWP, ZhuLL, ZhaoDY, LiHL. Microfluidic separation of particles by synergistic effect of geometry-induced hydrodynamics and magnetic field. Chin Chem Lett2024; 35(2):108646.
[7]
ZhangYS, AlemanJ, ShinSR, KilicT, KimD, MousaviShaegh SA, et al. Multisensor-integrated organs-on-chips platform for automated and continual in situ monitoring of organoid behaviors. Proc Natl Acad Sci USA2017; 114(12):E2293-302.
[8]
IsozakiA, HarmonJ, ZhouYQ, LiS, NakagawaY, HayashiM, et al. AI on a chip. Lab Chip2020; 20(17):3074-90.
[9]
JiaoYF, GaoLY, JiYR, LiuW. Recent advances in microfluidic single-cell analysis and its applications in drug development. Trends Analyt Chem2022;157:116796.
[10]
LiuJ, DuH, HuangL, XieW, LiuK, ZhangX, et al. AI-powered microfluidics: shaping the future of phenotypic drug discovery. ACS Appl Mater Interfaces2024; 16(30):38832-51.
[11]
Clausell-TormosJ, GriffithsAD, MertenCA. An automated two-phase microfluidic system for kinetic analyses and the screening of compound libraries. Lab Chip2010; 10(10):1302-7.
[12]
ZhaiJ, YiSH, JiaYW, MakPI, MartinsRP. Cell-based drug screening on microfluidics. Trends Analyt Chem2019;117:231-41.
[13]
KulesaA, KeheJ, HurtadoJE, TawdeP, BlaineyPC. Combinatorial drug discovery in nanoliter droplets. Proc Natl Acad Sci USA2018; 115 (26):6685-90.
[14]
AiYJ, ZhangF, WangCL, XieRX, LiangQL. Recent progress in lab-on-a-chip for pharmaceutical analysis and pharmacological/toxicological test. Trends Analyt Chem2019;117:215-30.
[15]
LowLA, MummeryC, BerridgeBR, AustinCP, TagleDA. Organs-on-chips: into the next decade. Nat Rev Drug Discov2021; 20(5):345-61.
[16]
SadeghzadeS, HooshiarMH, AkbariH, TajerMHM, SahnehKK, ZiaeiSY, et al. Recent advances in organ-on-a-chip models: how precision engineering integrates cutting edge technologies in fabrication and characterization. Appl Mater Today2024;38:102231.
[17]
LiuYX, SunLY, ZhangH, ShangLR, ZhaoYJ. Microfluidics for drug develop- ment: from synthesis to evaluation. Chem Rev2021; 121(13):7468-529.
LiJT, ChenJ, BaiH, WangHW, HaoSP, DingY, et al. An overview of organs-on- chips based on deep learning. Research2022;2022:9869518.
[21]
EkinsS, PuhlAC, ZornKM, LaneTR, RussoDP, KleinJJ, et al. Exploiting machine learning for end-to-end drug discovery and development. Nat Mater2019; 18(5):435-41.
[22]
ZhouJH, DongJP, HouHW, HuangL, LiJH. High-throughput microfluidic systems accelerated by artificial intelligence for biomedical applications. Lab Chip2024; 24(5):1307-26.
[23]
ZhengJH, ColeT, ZhangYX, KimJ, TangSY. Exploiting machine learning for bestowing intelligence to microfluidics. Biosens Bioelectron2021;194:113666.
[24]
PaulD, SanapG, ShenoyS, KalyaneD, KaliaK, TekadeRK. Artificial intelligence in drug discovery and development. Drug Discov Today2020; 26(1):80-93.
[25]
GormleyAJ. Machine learning in drug delivery. J Control Release2024;373:23-30.
[26]
BasileAO, YahiA, TatonettiNP. Artificial intelligence for drug toxicity and safety. Trends Pharmacol Sci2019; 40(9):624-35.
[27]
SharmaV, MottafeghA, JooJU, KangJH, WangL, KimDP. Toward microfluidic continuous-flow and intelligent downstream processing of biopharmaceuticals. Lab Chip2024; 24(11):2861-82.
[28]
WangJY, StevensJM, KariofillisSK, TomMJ, GoldenDL, LiJ, et al. Identifying general reaction conditions by bandit optimization. Nature2024; 626 (8001):1025-33.
[29]
ChenXY, LvHL. Intelligent control of nanoparticle synthesis on microfluidic chips with machine learning. NPG Asia Mater2022; 14(1):69.
[30]
YellenBB, ZawistowskiJS, CzechEA, SanfordCI, SoRelleED, LuftigMA, et al. Massively parallel quantification of phenotypic heterogeneity in single-cell drug responses. Sci Adv2021; 7(38):eabf9840.
[31]
BhhataraiB, WaltersWP, HopC, LanzaG, EkinsS. Opportunities and challenges using artificial intelligence in ADME/Tox. Nat Mater2019; 18 (5):418-22.
[32]
SchneiderG. Automating drug discovery. Nat Rev Drug Discov2018; 17 (2):97-113.
[33]
ReyesDR, IossifidisD, AurouxPA, ManzA. Micro total analysis systems. 1. Introduction, theory, and technology. Anal Chem2002; 74(12):2623-36.
[34]
CulbertsonCT, MickleburghTG, Stewart-JamesSA, SellensKA, PressnallM. Micro total analysis systems: fundamental advances and biological applications. Anal Chem2014; 86(1):95-118.
KulkarniJA, ThomsonSB, ZaifmanJ, LeungJ, WagnerPK, HillA, et al. Spontaneous, solvent-free entrapment of siRNA within lipid nanoparticles. Nanoscale2020; 12(47):23959-66.
[37]
ShepherdSJ, WarzechaCC, YadavaliS, El-MaytaR, AlamehMG, WangLL, et al. Scalable mRNA and siRNA lipid nanoparticle production using a parallelized microfluidic device. Nano Lett2021; 21(13):5671-80.
[38]
LiYJ, HuangXQ, LeeRJ, QiYH, WangKK, HaoF, et al. Synthesis of polymer- lipid nanoparticles by microfluidic focusing for siRNA delivery. Molecules2016; 21(10):1314.
[39]
ValenciaPM, FarokhzadOC, KarnikR, LangerR. Microfluidic technologies for accelerating the clinical translation of nanoparticles. Nat Nanotechnol2012; 7 (10):623-9.
[40]
JahnA, VreelandWN, GaitanM, LocascioLE. Controlled vesicle self-assembly in microfluidic channels with hydrodynamic focusing. J Am Chem Soc2004; 126(9):2674-5.
[41]
LuMQ, OzcelikA, GrigsbyCL, ZhaoYH, GuoF, LeongKW, et al. Microfluidic hydrodynamic focusing for synthesis of nanomaterials. Nano Today2016; 11 (6):778-92.
[42]
JahnA, StavisSM, HongJS, VreelandWN, DevoeDL, GaitanM. Microfluidic mixing and the formation of nanoscale lipid vesicles. ACS Nano2010; 4 (4):2077-87.
[43]
HoodRR, DeVoeDL. High-throughput continuous flow production of nanoscale liposomes by microfluidic vertical flow focusing. Small2015; 11 (43):5790-9.
[44]
LiuDF, BernuzCR, FanJ, LiW, CorreiaA, HirvonenJ, et al. A nano-in-nano vector: merging the best of polymeric nanoparticles and drug nanocrystals. Adv Funct Mater2017; 27(9):1604508.
[45]
YangJT, HuangKJ, LinYC. Geometric effects on fluid mixing in passive grooved micromixers. Lab Chip2005; 5(10):1140-7.
[46]
ChenDL, LoveKT, ChenY, EltoukhyAA, KastrupC, SahayG, et al. Rapid discovery of potent siRNA-containing lipid nanoparticles enabled by controlled microfluidic formulation. J Am Chem Soc2012; 134 (16):6948-51.
[47]
KastnerE, KaurR, LowryD, MoghaddamB, WilkinsonA, PerrieY. High- throughput manufacturing of size-tuned liposomes by a new microfluidics method using enhanced statistical tools for characterization. Int J Pharm2014; 477(1-2):361-8.
SzokaJr F, PapahadjopoulosD. Procedure for preparation of liposomes with large internal aqueous space and high capture by reverse-phase evaporation. Proc Natl Acad Sci USA1978; 75(9):4194-8.
MehrajiS, DeVoeDL. Microfluidic synthesis of lipid-based nanoparticles for drug delivery: recent advances and opportunities. Lab Chip2024; 24 (5):1154-74.
[52]
LiHX, WangXY, QiaoD, LiJH, ZhuWP, LiHL. Mechanism of nanoparticle aggregation in gas-liquid microfluidic mixing. Chin Chem Lett2024; 35 (4):108747.
[53]
SimonnetC, GroismanA. Chaotic mixing in a steady flow in a microchannel. Phys Rev Lett2005; 94(13):134501.
[54]
LiH, LiJ, QiaoD, WangX, ZhaoD, YanJ, et al. Mixing in a co-flow-focusing structured droplet-based micromixer. Chem Eng Sci2024;288:119854.
[55]
BorgognaA, MurmuraMA, AnnesiniMC, GionaM, CerbelliS. A hybrid numerical approach for predicting mixing length and mixing time in microfluidic junctions from moderate to arbitrarily large values of the Peclet number. Chem Eng Sci2019;196:247-64.
[56]
WörnerM. Approximate residence time distribution of fully develop laminar flow in a straight rectangular channel. Chem Eng Sci2010; 65(11):3499-507.
[57]
MariottiA, GallettiC, MauriR, SalvettiMV, BrunazziE. Effect of stratification on the mixing and reaction yield in a T-shaped micro-mixer. Phys Rev Fluids2021; 6(2):024202.
[58]
ShiYW, CaiY, CaoYH, HongZY, ChaiYF. Recent advances in microfluidic technology and applications for anti-cancer drug screening. Trends Analyt Chem2021;134:116118.
[59]
NguyenNT, ShaeghSAM, KashaninejadN, PhanDT. Design, fabrication and characterization of drug delivery systems based on lab-on-a-chip technology. Adv Drug Deliv Rev2013; 65(11-12):1403-19.
[60]
KhooBL, GrenciG, LimYB, LeeSC, HanJ, LimCT. Expansion of patient-derived circulating tumor cells from liquid biopsies using a CTC microfluidic culture device. Nat Protoc2018; 13(1):34-58.
[61]
FetahKL, DiPardoBJ, KongadzemEM, TomlinsonJS, ElzagheidA, ElmusratiM, et al. Cancer modeling-on-a-chip with future artificial intelligence integration. Small2019; 15(50):1901985.
[62]
ChungBG, FlanaganLA, RheeSW, SchwartzPH, LeeAP, MonukiES, et al. Human neural stem cell growth and differentiation in a gradient-generating microfluidic device. Lab Chip2005; 5(4):401-6.
[63]
IrimiaD, LiuSY, TharpWG, SamadaniA, TonerM, PoznanskyMC. Microfluidic system for measuring neutrophil migratory responses to fast switches of chemical gradients. Lab Chip2006; 6(2):191-8.
[64]
WuA, LoutherbackK, LambertG, Estévez-SalmerónL, TlstyTD, AustinRH, et al. Cell motility and drug gradients in the emergence of resistance to chemotherapy. Proc Natl Acad Sci USA2013; 110 (40):16103-8.
[65]
LeeWB, ChienCC, YouHL, KuoFC, LeeMS, LeeGB.An integrated microfluidic system for antimicrobial susceptibility testing with antibiotic combination. Lab Chip 2019; 19(16):2699-708.
[66]
KimJ, TaylorD, AgrawalN, WangH, KimH, HanA, et al. A programmable microfluidic cell array for combinatorial drug screening. Lab Chip2012; 12 (10):1813-22.
[67]
LiXJ, ChenYC, LiPCH. A simple and fast microfluidic approach of same- single-cell analysis (SASCA) for the study of multidrug resistance modulation in cancer cells. Lab Chip2011; 11(7):1378-84.
[68]
CavigliaC, ZórK, MontiniL, TilliV, CanepaS, MelanderF, et al. Impedimetric toxicity assay in microfluidics using free and liposome-encapsulated anticancer drugs. Anal Chem2015; 87(4):2204-12.
[69]
ZhuP, WangL. Passive and active droplet generation with microfluidics: a review. Lab Chip2017; 17(1):34-75.
[70]
NieZH, LiW, SeoM, XuSQ, KumachevaE. Janus and ternary particles generated by microfluidic synthesis: design, synthesis, and self-assembly. J Am Chem Soc2006; 128(29):9408-12.
[71]
AbateAR, WeitzDA. High-order multiple emulsions formed in poly(dimethylsiloxane) microfluidics. Small2009; 5(18):2030-2.
[72]
NabaviSA, VladisavljevicGT, ManovicV. Mechanisms and control of single- step microfluidic generation of multi-core double emulsion droplets. Chem Eng J2017;322:140-8.
[73]
ChoiCH, YiH, HwangS, WeitzDA, LeeCS. Microfluidic fabrication of complex-shaped microfibers by liquid template-aided multiphase microflow. Lab Chip2011; 11(8):1477-83.
WuJY, YadavaliS, LeeD, IssadoreDA. Scaling up the throughput of microfluidic droplet-based materials synthesis: a review of recent progress and outlook. Appl Phys Rev2021; 8(3):031304.
[79]
WangY, JinRN, ShenBQ, LiN, ZhouH, WangW, et al. High-throughput functional screening for next-generation cancer immunotherapy using droplet-based microfluidics. Sci Adv2021; 7(24):eabe3839.
QiaoD, LiHX, ChenXH, LiJH, WangXY, ZhaoDY, et al. Moving contact line dynamics for capillary-driven microfluidics in wetting transition regime. Phys Fluids2024; 36(5):052018.
[82]
UtadaAS, ChuLY, Fernandez-NievesA, LinkDR, HoltzeC, WeitzDA. Drippingjetting, drops, and wetting: the magic of microfluidics. MRS Bull2007; 32 (9):702-8.
[83]
NanL, ZhangHD, WeitzDA, ShumHC. Development and future of droplet microfluidics. Lab Chip2024; 24(5):1135-53.
[84]
LiHX, QiaoD, RenFZ, LiHL, ZhuWP, ZhaoDY, et al. Computational fluid dynamics analysis of droplet generation in microfluidic multi-cell coupled systems. Phys Fluids2021; 33(10):103313.
[85]
NisisakoT, ToriiT. Microfluidic large-scale integration on a chip for mass production of monodisperse droplets and particles. Lab Chip2008; 8 (2):287-93.
[86]
LiW, GreenerJ, VoicuD, KumachevaE. Multiple modular microfluidic (M3) reactors for the synthesis of polymer particles. Lab Chip 2009; 9(18):2715-21.
[87]
ConchousoD, CastroD, KhanSA, FouldsIG. Three-dimensional parallelization of microfluidic droplet generators for a litre per hour volume production of single emulsions. Lab Chip2014; 14(16):3011-20.
[88]
MulunehM, IssadoreD. Hybrid soft-lithography/laser machined microchips for the parallel generation of droplets. Lab Chip2013; 13(24):4750-4.
RomanowskyMB, AbateAR, RotemA, HoltzeC, WeitzDA. High throughput production of single core double emulsions in a parallelized microfluidic device. Lab Chip2012; 12(4):802-7.
[91]
YadavaliS, JeongHH, LeeD, IssadoreD. Silicon and glass very large scale microfluidic droplet integration for terascale generation of polymer microparticles. Nat Commun2018;9:1222.
[92]
JeongHH, ChenZ, YadavaliS, XuJH, IssadoreD, LeeD. Large-scale production of compound bubbles using parallelized microfluidics for efficient extraction of metal ions. Lab Chip2019; 19(4):665-73.
[93]
NawarS, StolaroffJK, YeCW, WuHY, NguyenDT, XinF, et al. Parallelizable microfluidic dropmakers with multilayer geometry for the generation of double emulsions. Lab Chip2020; 20(1):147-54.
MillerOJ, ElHarrak A, MangeatT, BaretJC, FrenzL, ElDebs B, et al. High- resolution dose-response screening using droplet-based microfluidics. Proc Natl Acad Sci USA2012; 109(2):378-83.
[99]
ToumaJG, CoblynM, FreibergLJ, KowallC, ZoebeleinA, JovanovicGN. Intensification of solvent extraction in an additively manufactured microfluidic separator. Chem Eng J2024;484:149285.
[100]
WangYT, ChenZY, BianFK, ShangLR, ZhuKX, ZhaoYJ. Advances of droplet- based microfluidics in drug discovery. Expert Opin Drug Discov2020; 15 (8):969-79.
SunilduttN, PariharP, SalihARC, LeeSH, ChoiKH. Revolutionizing drug development: harnessing the potential of organ-on-chip technology for disease modeling and drug discovery. Front Pharmacol2023;14:1139229.
[104]
Lorenzo-MartinLF, HubscherT, BowlerAD, BroguiereN, LangerJ, TillardL, et al. Spatiotemporally resolved colorectal oncogenesis in mini-colons ex vivo. Nature2024; 629(8011):450-7.
[105]
IngberDE. Human organs-on-chips for disease modelling, drug development and personalized medicine. Nat Rev Genet2022; 23(8):467-91.
[106]
HuhD, MatthewsBD, MammotoA, Montoya-ZavalaM, HsinHY, IngberDE. Reconstituting organ-level lung functions on a chip. Science2010; 328 (5986):1662-8.
[107]
KimHJ, HuhD, HamiltonG, IngberDE. Human gut-on-a-chip inhabited by microbial flora that experiences intestinal peristalsis-like motions and flow. Lab Chip2012; 12(12):2165-74.
[108]
BerthierE, YoungEWK, BeebeD. Engineers are from PDMS-land, biologists are from polystyrenia. Lab Chip2012; 12(7):1224-37.
[109]
ChanCY, GoralVN, DeRosaME, HuangTJ, YuenPK. A polystyrene-based microfluidic device with three-dimensional interconnected microporous walls for perfusion cell culture. Biomicrofluidics2014; 8(4):046505.
[110]
Olaizola-RodrigoC, Palma-FlorezS, RandelovicT, BayonaC, AshrafiM, SamitierJ, et al. Tuneable hydrogel patterns in pillarless microfluidic devices. Lab Chip2024; 24(7):2094-106.
[111]
KoleskyDB, HomanKA, Skylar-ScottMA, LewisJA.Three-dimensional bioprinting of thick vascularized tissues. Proc Natl Acad Sci USA 2016; 113 (12):3179-84.
[112]
LiuHT, WangYQ, CuiKL, GuoYQ, ZhangX, QinJH. Advances in hydrogels in organoids and organs-on-a-chip. Adv Mater2019; 31(50):1902042.
[113]
CampbellSB, WuQH, YazbeckJ, LiuCA, OkhovatianS, RadisicM. Beyond polydimethylsiloxane: alternative materials for fabrication of organ-on-a- chip devices and microphysiological systems. ACS Biomater Sci Eng2021; 7 (7):2880-99.
[114]
Juste-LanasY, Hervas-RaluyS, Garcia-AznarJM, Gonzalez-LoyolaA.Fluid flow to mimic organ function in 3D in vitro models. APL Bioeng 2023; 7 (3):031501.
[115]
AgbesiRJA, ChevalierNR. Flow and mixing induced by single, colinear, and colliding contractile waves in the intestine. Phys Rev Fluids2022; 7 (4):024202.
[116]
BenamKH, VillenaveR, LucchesiC, VaroneA, HubeauC, LeeHH, et al. Small airway-on-a-chip enables analysis of human lung inflammation and drug responses in vitro. Nat Methods2016; 13(2):151-7.
[117]
KimHJ, IngberDE. Gut-on-a-chip microenvironment induces human intestinal cells to undergo villus differentiation. Integr Biol2013; 5 (9):1130-40.
[118]
Corral-NájeraK, ChauhanG, Serna-SaldívarSO, Martínez-ChapaSO, AeinehvandM. Polymeric and biological membranes for organ-on-a-chip devices. Microsyst Nanoeng2023; 9(1):107.
[119]
HanY, ZhouZ, JinR, DaiF, GeY, JuX, et al. Mechanical activation opens a lipid-lined pore in OSCA ion channels. Nature2024; 628(8009):910-8.
[120]
QuintardC, TubbsE, JonssonG, JiaoJ, WangJ, WerschlerN, et al. A microfluidic platform integrating functional vascularized organoids-on-chip. Nat Commun2024;15:1452.
[121]
ChenP, LiSJ, GuoYR, ZengXM, LiuBF. A review on microfluidics manipulation of the extracellular chemical microenvironment and its emerging application to cell analysis. Anal Chim Acta2020;1125:94-113.
ZhaoYM, KoroljA, FericN, RadisicM. Human pluripotent stem cell-derived cardiomyocyte based models for cardiotoxicity and drug discovery. Expert Opin Drug Saf2016; 15(11):1455-8.
[124]
AhadianS, CivitareseR, BannermanD, MohammadiMH, LuR, WangE, et al. Organ-on-a-chip platforms: a convergence of advanced materials, cells, and microscale technologies. Adv Healthc Mater2018; 7(2):1700506.
[125]
KolanowskiTJ, BusekM, SchubertM, DmitrievaA, BinnewergB, PöcheJ, et al. Enhanced structural maturation of human induced pluripotent stem cell- derived cardiomyocytes under a controlled microenvironment in a microfluidic system. Acta Biomater2020;102:273-86.
[126]
VatineGD, BarrileR, WorkmanMJ, SancesS, BarrigaBK, RahnamaM, et al. Human iPSc-derived blood-brain barrier chips enable disease modeling and personalized medicine applications. Cell Stem Cell2019; 24(6):995-1005.
[127]
FangGC, LuHX, Al-NakashliR, ChapmanR, ZhangYQ, JuLA, et al. Enabling peristalsis of human colon tumor organoids on microfluidic chips. Biofabrication2022; 14(1):015006.
[128]
WuL, AiYJ, XieRX, XiongJL, WangY, LiangQL. Organoids/organs-on-a-chip: new frontiers of intestinal pathophysiological models. Lab Chip2023; 23 (5):1192-212.
[129]
BaiHQ, SiLL, JiangA, BelgurC, ZhaiYH, PlebaniR, et al. Mechanical control of innate immune responses against viral infection revealed in a human lung alveolus chip. Nat Commun2022;13:1928.
[130]
MaC, PengYS, LiHT, ChenWQ. Organ-on-a-chip: a new paradigm for drug development. Trends Pharmacol Sci2021; 42(2):119-33.
[131]
PereiraSAP, DysonPJ, SaraivaM. Miniaturized technologies for high- throughput drug screening enzymatic assays and diagnostics—a review. Trends Analyt Chem2020;126:115862.
[132]
BaudoinR, ProtJM, NicolasG, BrochetonJ, BrochotC, LegallaisC, et al. Evaluation of seven drug metabolisms and clearances by cryopreserved human primary hepatocytes cultivated in microfluidic biochips. Xenobiotica2013; 43(2):140-52.
[133]
FuJY, QiuHL, TanCS. Microfluidic liver-on-a-chip for preclinical drug discovery. Pharmaceutics2023; 15(4):1300.
[134]
VidiPA, MalekiT, OchoaM, WangL, ClarkSM, LearyJF, et al. Disease-on-a- chip: mimicry of tumor growth in mammary ducts. Lab Chip2014; 14 (1):172-7.
[135]
TatosianDA, ShulerML. A novel system for evaluation of drug mixtures for potential efficacy in treating multidrug resistant cancers. Biotechnol Bioeng2009; 103(1):187-98.
[136]
KhetaniSR, BhatiaSN. Microscale culture of human liver cells for drug development. Nat Biotechnol2008; 26(1):120-6.
[137]
GoughA, Soto-GutierrezA, VernettiL, EbrahimkhaniMR, SternAM, TaylorDL. Human biomimetic liver microphysiology systems in drug development and precision medicine. Nat Rev Gastroenterol Hepatol2021; 18(4):252-68.
[138]
JungDJ, ShinTH, KimM, SungCO, JangSJ, JeongGS. A one-stop microfluidic- based lung cancer organoid culture platform for testing drug sensitivity. Lab Chip2019; 19(17):2854-65.
[139]
MaschmeyerI, LorenzAK, SchimekK, HasenbergT, RammeAP, HübnerJ, et al. A four-organ-chip for interconnected long-term co-culture of human intestine, liver, skin and kidney equivalents. Lab Chip2015; 15(12):2688-99.
[140]
SahaB, DasB, MajumderM. A deep-reinforcement learning approach for optimizing homogeneous droplet routing in digital microfluidic biochips. Nanotechnol Precis Eng2023; 6(2):023001.
[141]
KoJ, BhagwatN, YeeSS, OrtizN, SahmoudA, BlackT, et al. Combining machine learning and nanofluidic technology to diagnose pancreatic cancer using exosomes. ACS Nano2017; 11(11):11182-93.
[142]
HonradoC, BisegnaP, SwamiNS, CaselliF. Single-cell microfluidic impedance cytometry: from raw signals to cell phenotypes using data analytics. Lab Chip2021; 21(1):22-54.
[143]
EllettF, JorgensenJ, MarandAL, LiuYM, MartinezMM, SeinV, et al. Diagnosis of sepsis from a drop of blood by measurement of spontaneous neutrophil motility in a microfluidic assay. Nat Biomed Eng2018; 2(4):207-14.
[144]
ShippMA, RossKN, TamayoP, WengAP, KutokJL, AguiarRCT, et al. Diffuse large B-cell lymphoma outcome prediction by gene-expression profiling and supervised machine learning. Nat Med2002; 8(1):68-74.
[145]
HonradoC, SalahiA, AdairSJ, MooreJH, BauerTW, SwamiNS. Automated biophysical classification of apoptotic pancreatic cancer cell subpopulations by using machine learning approaches with impedance cytometry. Lab Chip2022; 22(19):3708-20.
[146]
FilippiJ, DiGiuseppe D, CastiP, MencattiniA, AntonelliG, D’OrazioM, et al. Exploiting spectral information in opto-electronic tweezers for cell classification and drug response evaluation. Sensor Actuat Biol Chem2022;368:132200.
[147]
MonzelAS, HemmerK, KaomaT, SmitsLM, BologninS, LucarelliP, et al. Machine learning-assisted neurotoxicity prediction in human midbrain organoids. Parkinsonism Relat Disord2020;75:105-9.
[148]
LuMJ, RaoSL, YueH, HanJJ, WangJT. Recent advances in the application of machine learning to crystal behavior and crystallization process control. Cryst Growth Des2024; 24(12):5374-96.
[149]
SatoS, SanoS, MutoH, KubaraK, KondoK, MiyazakiT, et al. Understanding the manufacturing process of lipid nanoparticles for mRNA delivery using machine learning. Chem Pharm Bull2024; 72(6):529-39.
[150]
van de StolpeA, KauffmannRH. Innovative human-specific investigational approaches to autoimmune disease. RSC Adv2015; 5(24):18451-63.
[151]
LeCunY, BengioY, HintonG. Deep learning. Nature2015; 521(7553): 436-44.
[152]
GrisoniF, HuismanBJH, ButtonAL, MoretM, AtzK, MerkD, et al. Combining generative artificial intelligence and on-chip synthesis for de novo drug design. Sci Adv2021; 7(24):eabg3338.
[153]
DeanSN, AlvarezJAE, ZabetakisD, WalperSA, MalanoskiAP. PepVAE: variational autoencoder framework for antimicrobial peptide generation and activity prediction. Front Microbiol2021;12:725727.
[154]
RudinC. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead. Nat Mach Intell2019; 1 (5):206-15.
[155]
SayersEW, BoltonEE, BristerJR, CaneseK, ChanJ, ComeauDC, et al. Database resources of the national center for biotechnology information in 2023. Nucleic Acids Res2023; 51(D1):D29-38.
[156]
AhmadiF, SimchiM, PerryJM, FrenetteS, BenaliH, SoucyJP, et al. Integrating machine learning and digital microfluidics for screening experimental conditions. Lab Chip2022; 23(1):81-91.
[157]
CaicedoJC, CooperS, HeigwerF, WarchalS, QiuP, MolnarC, et al. Data- analysis strategies for image-based cell profiling. Nat Methods2017; 14 (9):849-63.
[158]
McIntyreD, LashkaripourA, FordyceP, DensmoreD. Machine learning for microfluidic design and control. Lab Chip2022; 22(16):2925-37.
[159]
LiuPW, JinH, ChenY, WangDR, YanHH, WuMZ, et al. Process analytical technologies and self-optimization algorithms in automated pharmaceutical continuous manufacturing. Chin Chem Lett2024; 35(3):108877.
[160]
MeissnerF, Geddes-McAlisterJ, MannM, BantscheffM. The emerging role of mass spectrometry-based proteomics in drug discovery. Nat Rev Drug Discov2022; 21(9):637-54.
[161]
GebreyesusST, MuneerG, HuangCC, SiyalAA, AnandM, ChenYJ, et al. Recent advances in microfluidics for single-cell functional proteomics. Lab Chip2023; 23(7):1726-51.
[162]
GebreyesusST, SiyalAA, KitataRB, ChenESW, EnkhbayarB, AngataT, et al. Streamlined single-cell proteomics by an integrated microfluidic chip and data-independent acquisition mass spectrometry. Nat Commun2022;13:37.
[163]
LiuCC, XuXN, LiB, SituB, PanWL, HuY, et al. Single-exosome-counting immunoassays for cancer diagnostics. Nano Lett2018; 18(7):4226-32.
[164]
XieJX, RuanSS, TuMY, YuanZ, HuJG, LiHL, et al. Clustering single-cell RNA sequencing data via iterative smoothing and self-supervised discriminative embedding. Oncogene2024; 43(29):2279-92.
[165]
GrahamG, CsicseryN, StasiowskiE, ThouveninG, MatherWH, FerryM, et al. Genome-scale transcriptional dynamics and environmental biosensing. Proc Natl Acad Sci USA2020; 117(6):3301-6.
[166]
PiazzaI, BeatonN, BrudererR, KnoblochT, BarbisanC, ChandatL, et al. A machine learning-based chemoproteomic approach to identify drug targets and binding sites in complex proteomes. Nat Commun2020;11:4200.
[167]
MigliozziS, OhYT, HasanainM, GarofanoL, D’AngeloF, NajacRD, et al. Integrative multi-omics networks identify PKCd and DNA-PK as master kinases of glioblastoma subtypes and guide targeted cancer therapy. Nat Can2023; 4(2):181-202.
[168]
JeonJ, NimS, TeyraJ, DattiA, WranaJL, SidhuSS, et al. A systematic approach to identify novel cancer drug targets using machine learning, inhibitor design and high-throughput screening. Genome Med2014; 6(7):6.
[169]
EschEW, BahinskiA, HuhD. Organs-on-chips at the frontiers of drug discovery. Nat Rev Drug Discov2015; 14(4):248-60.
[170]
BersiniS, JeonJS, DubiniG, ArrigoniC, ChungS, CharestJL, et al. A microfluidic 3D in vitro model for specificity of breast cancer metastasis to bone. Biomaterials2014; 35(8):2454-61.
[171]
DeChiara F, Ferret-MiñanaA, Ramón-AzcónJ. The synergy between organ- on-a-chip and artificial intelligence for the study of NAFLD: from basic science to clinical research. Biomedicines2021; 9(3):248.
[172]
BuggenthinF, BuettnerF, HoppePS, EndeleM, KroissM, StrasserM, et al. Prospective identification of hematopoietic lineage choice by deep learning. Nat Methods2017; 14(4):403-6.
[173]
DengLY, HuangXY, RenJC. In situ study of the drug-target protein interaction in single living cells by combining fluorescence correlation spectroscopy with affinity probes. Anal Chem2020; 92(10):7020-7.
[174]
D’AmicoCI, PolaskyDA, SteyerDJ, RuotoloBT, KennedyRT. Ion mobility-mass spectrometry coupled to droplet microfluidics for rapid protein structure analysis and drug discovery. Anal Chem2022; 94(38):13084-91.
[175]
LiuZB, HanX, ChenR, ZhangK, LiY, FrugeS, et al. Microfluidic mapping of cancer cell-protein binding interaction. ACS Appl Mater Interfaces2017; 9 (27):22143-8.
[176]
DingYH, BallKA, WebbKJ, GaoY, D’AlessandroA, OldWM, et al. On-chip acousto thermal shift assay for rapid and sensitive assessment of protein thermodynamic stability. Small2020; 16(41):2003506.
[177]
SanicasM, TorroR, LimozinL, ChamesP. Antigen density and applied force control enrichment of nanobody-expressing yeast cells in microfluidics. Lab Chip2024; 24(11):2944-57.
[178]
WalgamaC, AlMubarak ZH, ZhangB, AkinwaleM, PathiranageA, DengJP, et al. Label-free real-time microarray imaging of cancer protein-protein interactions and their inhibition by small molecules. Anal Chem2016; 88 (6):3130-5.
[179]
WangWZ, WeiZW, ZhangD, MaHL, WangZH, BuXL, et al. Rapid screening of peptide probes through in situ single-bead sequencing microarray. Anal Chem2014; 86(23):11854-9.
SinghR, SledzieskiS, BrysonB, CowenL, BergerB. Contrastive learning in protein language space predicts interactions between drugs and protein targets. Proc Natl Acad Sci USA2023; 120(24):e2220778120.
[183]
KarimiM, WuD, WangZY, ShenY. DeepAffinity: interpretable deep learning of compound-protein affinity through unified recurrent and convolutional neural networks. Bioinformatics2019; 35(18):3329-38.
SharmaA, RaniR. BE-DTI′: ensemble framework for drug target interaction prediction using dimensionality reduction and active learning. Comput Methods Programs Biomed2018;165:151-62.
[186]
YuWB, WeberDJ, MacKerellJr AD. Integrated covalent drug design workflow using site identification by ligand competitive saturation. J Chem Theory Comput2023; 19(10):3007-21.
[187]
ZhouJW, EllisAV, KobusH, VoelckerNH. Aptamer sensor for cocaine using minor groove binder based energy transfer. Anal Chim Acta2012;719:76-81.
[188]
MollaeiP, FarimaniAB. Activity map and transition pathways of G protein- coupled receptor revealed by machine learning. J Chem Inf Model2023; 63 (8):2296-304.
[189]
FengJG, NeuzilJ, ManzA, IliescuC, NeuzilP. Microfluidic trends in drug screening and drug delivery. Trends Analyt Chem2023;158:116821.
[190]
Jiménez-LunaJ, GrisoniF, WeskampN, SchneiderG. Artificial intelligence in drug discovery: recent advances and future perspectives. Expert Opin Drug Discov2021; 16(9):949-59.
[191]
QingLS, WangTT, LuoHY, DuJL, WangRY, LuoP. Microfluidic strategies for natural products in drug discovery: current status and future perspectives. Trends Analyt Chem2023;158:116832.
[192]
SunJH, WardenAR, DingXT. Recent advances in microfluidics for drug screening. Biomicrofluidics2019; 13(6):061503.
[193]
HorowitzLF, RodriguezAD, Au-YeungA, BishopKW, BarnerLA, MishraG, et al. Microdissected “cuboids” for microfluidic drug testing of intact tissues. Lab Chip2021; 21(1):122-42.
[194]
delRío JS, RoJ, YoonH, ParkTE, ChoYK. Integrated technologies for continuous monitoring of organs-on-chips: current challenges and potential solutions. Biosens Bioelectron2023;224:115057.
[195]
ChanKK, ShangLW, QiaoZ, LiaoYK, KimM, ChenYC.Monitoring amyloidogenesis with a 3D deep-learning-guided biolaser imaging array. Nano Lett 2022; 22(22):8949-56.
[196]
LinXD, DuanX, JacobsC, UllmannJ, ChanCY, ChenSY, et al. High-throughput brain activity mapping and machine learning as a foundation for systems neuropharmacology. Nat Commun2018;9:5142.
[197]
YuJ, LeeS, SongJ, LeeSR, KimS, ChoiH, et al. Perfusable micro-vascularized 3D tissue array for high-throughput vascular phenotypic screening. Nano Converg2022; 9(1):16.
[198]
WangG, McCainML, YangLH, HeAB, PasqualiniFS, AgarwalA, et al. Modeling the mitochondrial cardiomyopathy of Barth syndrome with induced pluripotent stem cell and heart-on-chip technologies. Nat Med2014; 20(6):616-23.
[199]
AoZ, CaiHW, WuZH, HuL, NunezA, ZhouZL, et al. Microfluidics guided by deep learning for cancer immunotherapy screening. Proc Natl Acad Sci USA2022; 119(46):e2214569119.
[200]
AvilaC, CassaniC, KogejT, MazuelaJ, SardaS, ClaytonAD, et al. Automated stopped-flow library synthesis for rapid optimisation and machine learning directed experimentation. Chem Sci2022; 13(41):12087-99.
[201]
ScottHL, BucknerN, Fernandez-AlbertF, PedoneE, PostiglioneL, ShiGY, et al. A dual druggable genome-wide siRNA and compound library screening approach identifies modulators of parkin recruitment to mitochondria. J Biol Chem2020; 295(10):3285-300.
[202]
GryniukovaA, KaiserF, MyziukI, AlieksieievaD, LeberechtC, HeymPP, et al. AI-powered virtual screening of large compound libraries leads to the discovery of novel inhibitors of Sirtuin-1. J Med Chem 2023; 66 (15):10241-51.
[203]
XieCL, ZhuangXX, NiuZM, AiRX, LautrupS, ZhengSJ, et al. Amelioration of Alzheimer’s disease pathology by mitophagy inducers identified via machine learning and a cross-species workflow. Nat Biomed Eng2022; 6(1):76-93.
[204]
WangC, WuYH, XueYF, ZouLY, HuangY, ZhangP, et al. Combinatorial discovery of antibacterials via a feature-fusion based machine learning workflow. Chem Sci2024; 15(16):6044-52.
[205]
ManakMS, VarsanikJS, HoganBJ, WhitfieldMJ, SuWR, JoshiN, et al. Live-cell phenotypic-biomarker microfluidic assay for the risk stratification of cancer patients via machine learning. Nat Biomed Eng2018; 2(10):761-72.
[206]
ChongLH, ChingT, FarmHJ, GrenciG, ChiamKH, TohYC. Integration of a microfluidic multicellular coculture array with machine learning analysis to predict adverse cutaneous drug reactions. Lab Chip2022; 22(10):1890-904.
[207]
PreuerK, LewisRPI, HochreiterS, BenderA, BulusuKC, KlambauerG. DeepSynergy: predicting anti-cancer drug synergy with deep learning. Bioinformatics2018; 34(9):1538-46.
[208]
JulkunenH, CichonskaA, GautamP, SzedmakS, DouatJ, PahikkalaT, et al. Leveraging multi-way interactions for systematic prediction of pre-clinical drug combination effects. Nat Commun2020;11:6136.
[209]
JanizekJD, DincerAB, CelikS, ChenHG, ChenWL, NaxerovaK, et al. Uncovering expression signatures of synergistic drug responses via ensembles of explainable machine-learning models. Nat Biomed Eng2023; 7(6):811-29.
[210]
IanevskiA, GiriAK, GautamP, KononovA, PotdarS, SaarelaJ, et al. Prediction of drug combination effects with a minimal set of experiments. Nat Mach Intell2019; 1(12):568-77.
[211]
MalyutinaA, MajumderMM, WangWY, PessiaA, HeckmanCA, TangJ. Drug combination sensitivity scoring facilitates the discovery of synergistic and efficacious drug combinations in cancer. PLoS Comput Biol2019; 15(5): e1006752.
[212]
SuHR, MaTX, LiuX, WangL, ShuFJ, LiangZQ, et al. Microfluidic organ chip of fluid-solid dynamic curved interface. Appl Phys Rev2024; 11(1):011404.
[213]
MaoYL, WickströmSA. Mechanical state transitions in the regulation of tissue form and function. Nat Rev Mol Cell Biol2024; 25(8):654-70.
[214]
FangY, EglenRM. Three-dimensional cell cultures in drug discovery and development. SLAS Discov2017; 22(5):456-72.
[215]
ParkTE, MustafaogluN, HerlandA, HasselkusR, MannixR, FitzGeraldEA, et al. Hypoxia-enhanced blood-brain barrier chip recapitulates human barrier function and shuttling of drugs and antibodies. Nat Commun2019;10:2621.
[216]
LeeDH, TranHQ, DudleyAT, YangK, YanZ, XieJW. Advancing nerve regeneration: peripheral nerve injury (PNI) chip empowering high-speed biomaterial and drug screening. Chem Eng J2024;486:150210.
[217]
LiuYJ, ZhouYB, ChenP. Lung cancer organoids: models for preclinical research and precision medicine. Front Oncol2023;13:1293441.
[218]
ChenZK, KheiriS, GevorkianA, YoungEWK, AndreV, DeisenrothT, et al. Microfluidic arrays of dermal spheroids: a screening platform for active ingredients of skincare products. Lab Chip2021; 21(20):3952-62.
[219]
BlackBJ, AtmaramaniR, PlagensS, CampbellZT, DussorG, PriceTJ, et al. Emerging neurotechnology for antinoceptive mechanisms and therapeutics discovery. Biosens Bioelectron2019;126:679-89.
[220]
FangGC, LuHX, de la FuenteLR, LawAMK, LinGG, JinDY, et al. Mammary tumor organoid culture in non-adhesive alginate for luminal mechanics and high-throughput drug screening. Adv Sci2021; 8(21):2102418.
[221]
LeeRY, WuY, GohD, TanV, NgCW, LimJCT, et al. Application of artificial intelligence to in vitro tumor modeling and characterization of the tumor microenvironment. Adv Healthc Mater2023; 12(14):2202457.
[222]
PhamTH, QiuY, ZengJC, XieL, ZhangP.A deep learning framework for high- throughput mechanism-driven phenotype compound screening and its application to COVID-19 drug repurposing. Nat Mach Intell 2021; 3 (3):247-57.
[223]
KusumotoD, SekiT, SawadaH, KunitomiA, KatsukiT, KimuraM, et al. Anti- senescent drug screening by deep learning-based morphology senescence scoring. Nat Commun2021;12:257.
[224]
KrentzelD, ShorteSL, ZimmerC. Deep learning in image-based phenotypic drug discovery. Trends Cell Biol2023; 33(7):538-54.
[225]
YipJK, HarrisonM, VillafuerteJ, FernandezGE, PetersenAP, LienCL, et al. Extended culture and imaging of normal and regenerating adult zebrafish hearts in a fluidic device. Lab Chip2020; 20(2):274-84.
[226]
LockerySR, HulmeSE, RobertsWM, RobinsonKJ, LaromaineA, LindsayTH, et al. A microfluidic device for whole-animal drug screening using electrophysiological measures in the nematode C. elegans. Lab Chip2012; 12(12):2211-20.
[227]
TangMH, DuanX, YangAQ, HeSJ, ZhouYJ, LiuYX, et al. Fish capsules: a system for high-throughput screening of combinatorial drugs. Adv Sci2022; 9 (9):2104449.
[228]
LiuCT, WangYC, ZengYX, KangZR, ZhaoH, QiK, et al. Use of deep-learning assisted assessment of cardiac parameters in zebrafish to discover cyanidin chloride as a novel Keap 1 inhibitor against doxorubicin-induced cardiotoxicity. Adv Sci2023; 10(30):2301136.
[229]
XuJG, HuangMS, WangHF, FangQ. Forming a large-scale droplet array in a microcage array chip for high-throughput screening. Anal Chem2019; 91 (16):10757-63.
[230]
Holland-MoritzDA, WismerMK, MannBF, FarasatI, DevineP, GuetschowED, et al. Mass activated droplet sorting (MADS) enables high-throughput screening of enzymatic reactions at nanoliter scale. Angew Chem Int Ed2020; 59(11):4470-7.
[231]
JuskováP, SchmidYRF, StuckiA, SchmittS, HeldM, DittrichPS. “Basicles”: microbial growth and production monitoring in giant lipid vesicles. ACS Appl Mater Interfaces2019; 11(38):34698-706.
[232]
ChurskiK, KaminskiTS, JakielaS, KamyszW, Baranska-RybakW, WeibelDB, et al. Rapid screening of antibiotic toxicity in an automated microdroplet system. Lab Chip 2012; 12(9):1629-37.
[233]
MaFQ, ChungMT, YaoY, NidetzR, LeeLM, LiuAP, et al. Efficient molecular evolution to generate enantioselective enzymes using a dual-channel microfluidic droplet screening platform. Nat Commun2018;9:1030.
[234]
TanZJ, YanYW, LiaoJC, ShiHH, ZhengY, XuWZ, et al. Droplet array with microfluidic concentration gradient (DA-MCG) for 2-dimensional reaction condition screening. Chem Eng Sci2024;298:120432.
[235]
SarkarS, CohenN, SabhachandaniP, KonryT. Phenotypic drug profiling in droplet microfluidics for better targeting of drug-resistant tumors. Lab Chip2015; 15(23):4441-50.
[236]
DuGS, PanJZ, ZhaoSP, ZhuY, denToonder JMJ, FangQ. Cell-based drug combination screening with a microfluidic droplet array system. Anal Chem2013; 85(14):6740-7.
[237]
SesenM, AlanT, NeildA. Droplet control technologies for microfluidic high throughput screening (lHTS). Lab Chip2017; 17(14):2372-94.
[238]
ReutlingerM, RodriguesT, SchneiderP, SchneiderG. Combining on-chip synthesis of a focused combinatorial library with computational target prediction reveals imidazopyridine GPCR ligands. Angew Chem Int Ed2014; 53(2):582-5.
MacConnellAB, PriceAK, PaegelBM. An integrated microfluidic processor for DNA-encoded combinatorial library functional screening. ACS Comb Sci2017; 19(3):181-92.
[241]
YangZH, ZhouZH, SiTX, ZhouZD, ZhouL, ChinYR, et al. High throughput confined migration microfluidic device for drug screening. Small2023; 19 (16):2207194.
[242]
DidierCM, OrricoJF, TorresOCS, CastroJM, BakshA, RajaramanS. Microfabricated polymer-metal biosensors for multifarious data collection from electrogenic cellular models. Microsyst Nanoeng2023; 9(1):22.
[243]
HouY, YaoHR, LinJM. Recent advancements in single-cell metabolic analysis for pharmacological research. J Pharm Anal2023; 13(10):1102-16.
[244]
YangCG, WuYF, XuZR, WangJH. A radial microfluidic concentration gradient generator with high-density channels for cell apoptosis assay. Lab Chip2011; 11(19):3305-12.
[245]
ZhengXT, YuL, LiPW, DongH, WangYJ, LiuY, et al. On-chip investigation of cell-drug interactions. Adv Drug Deliv Rev2013; 65(11-12):1556-74.
[246]
PostekW, GarsteckiP. Droplet microfluidics for high-throughput analysis of antibiotic susceptibility in bacterial cells and populations. Acc Chem Res2022; 55(5):605-15.
[247]
YeNN, QinJH, ShiWW, LiuX, LinBC. Cell-based high content screening using an integrated microfluidic device. Lab Chip2007; 7(12):1696-704.
[248]
ZhangF, TianC, LiuWM, WangK, WeiYQ, WangHS, et al. Determination of benzopyrene-induced lung inflammatory and cytotoxic injury in a chemical gradient-integrated microfluidic bronchial epithelium system. ACS Sens2018; 3(12):2716-25.
[249]
AiXN, WuY, LuWB, ZhangXR, ZhaoL, TuPF, et al. A precise microfluidic assay in single-cell profile for screening of transient receptor potential channel modulators. Adv Sci2020; 7(11):2000111.
[250]
JiangY, RenXL, LiuGB, ChenSL, HaoM, DengXR, et al. Exploring the mechanism of contact-dependent cell-cell communication on chemosensitivity based on single-cell high-throughput drug screening platform. Talanta 2024;273:125869.
ZhaoSK, HuXJ, ZhuJM, LuoZY, LiangL, YangDY, et al. On-chip rapid drug screening of leukemia cells by acoustic streaming. Lab Chip2021; 21 (20):4005-15.
[254]
ZhaoMW, YangJ, LiZQ, ZengY, TaoCX, DaiB, et al. High-throughput 3D microfluidic chip for generation of concentration gradients and mixture combinations. Lab Chip2024; 24(8):2280-6.
MullardA. Microfluidics platform lowers barrier to drug combination screening. Nat Rev Drug Discov2018; 17(10):691-2.
[257]
DaiB, LongY, WuJD, HuangSQ, ZhaoY, ZhengLL, et al. Generation of flow and droplets with an ultra-long-range linear concentration gradient. Lab Chip2021; 21(22):4390-400.
[258]
SamimiA, HengojuS, RosenbaumMA. Combinatorial sample preparation platform for droplet-based applications in microbiology. Sens Actuators B2024;417:136162.
[259]
EduatiF, UtharalaR, MadhavanD, NeumannUP, LongerichT, CramerT, et al. A microfluidics platform for combinatorial drug screening on cancer biopsies. Nat Commun2018;9:2434.
[260]
UtharalaR, GrabA, VafaizadehV, PeschkeN, BallingerM, TureiD, et al. A microfluidic Braille valve platform for on-demand production, combinatorial screening and sorting of chemically distinct droplets. Nat Protoc2022; 17 (12):2920-65.
[261]
SteinbergE, FriedmanR, GoldsteinY, FriedmanN, BeharierO, DemmaJA, et al. A fully 3D-printed versatile tumor-on-a-chip allows multi-drug screening and correlation with clinical outcomes for personalized medicine. Commun Biol2023; 6(1):1157.
[262]
ChangHC, LinCH, JuangDE, WuHW, LeeCY, ChenCC, et al. Multilayer architecture microfluidic network array for combinatorial drug testing on 3D-cultured cells. Biofabrication2019; 11(3):035024.
[263]
DaviesM, AbubakerM, BibleLA. Flexiblemicrofluidic, dispensing system for screening drug combinations. Micromachines2020; 11(10):943.
[264]
WaringMJ, ArrowsmithJ, LeachAR, LeesonPD, MandrellS, OwenRM, et al. An analysis of the attrition of drug candidates from four major pharmaceutical companies. Nat Rev Drug Discov2015; 14(7):475-86.
[265]
RommEL, TsigelnyIF. Artificial intelligence in drug treatment. Ann Rev Pharmacol Toxicol2020;60:353-69.
[266]
ChengFX, LiWH, LiuGX, TangY. In silico ADMET prediction: recent advances, current challenges and future trends. Curr Top Med Chem2013; 13 (11):1273-89.
[267]
FerreiraLLG, AndricopuloAD. ADMET modeling approaches in drug discovery. Drug Discov Today2019; 24(5):1157-65.
[268]
ChenEP, BondiRW, MichalskiPJ.Model-based target pharmacology assessment (mTPA): an approach using PBPK/PD modeling and machine learning to design medicinal chemistry and DMPK strategies in early drug discovery. J Med Chem 2021; 64(6):3185-96.
[269]
ChouWC, ChenQR, YuanL, ChengYH, HeCL, Monteiro-RiviereNA, et al. An artificial intelligence-assisted physiologically-based pharmacokinetic model to predict nanoparticle delivery to tumors in mice. J Control Release2023;361:36153-63.
[270]
LuJM, BenderB, JinJY, GuanYF. Deep learning prediction of patient response time course from early data via neural-pharmacokinetic/pharmacodynamic modelling. Nat Mach Intell2021; 3(8):696-704.
[271]
PeteaniG, HuynhMTD, GerebtzoffG, Rodríguez-PérezR. Application of machine learning models for property prediction to targeted protein degraders. Nat Commun2024;15:5764.
[272]
FengCL, ChenHW, YuanXQ, SunMQ, ChuKX, LiuHQ, et al. Gene expression data based deep learning model for accurate prediction of drug-induced liver injury in advance. J Chem Inf Model2019; 59(7):3240-50.
[273]
MaHH, AnWZ, WangYH, SunHM, HuangRL, HuangJZ. Deep graph learning with property augmentation for predicting drug-induced liver injury. Chem Res Toxicol2021; 34(2):495-506.
[274]
ChenX, RobertsR, LiuZC, TongWD. A generative adversarial network model alternative to animal studies for clinical pathology assessment. Nat Commun2023;14:7141.
[275]
KandasamyK, ChuahJKC, SuR, HuangP, EngKG, XiongSJ, et al. Prediction of drug-induced nephrotoxicity and injury mechanisms with human induced pluripotent stem cell-derived cells and machine learning methods. Sci Rep2015; 5(1):12337.
[276]
ZhuZJ, WuRC, LuoM, ZengLH, ZhangDM, HuN, et al. Two-dimensional deep learning frameworks for drug-induced cardiotoxicity detection. ACS Sens2024; 9(6):3316-26.
[277]
WangYQ, WangL, ZhuYJ, QinJH. Human brain organoid-on-a-chip to model prenatal nicotine exposure. Lab Chip2018; 18(6):851-60.
[278]
Jimenez-CarreteroD, AbrishamiV, Fernández-de-ManuelL, PalaciosI, Quilez-Alvarez A, Díez-SánchezA, et al. Tox_(R)CNN: deep learning-based nuclei profiling tool for drug toxicity screening. PLoS Comput Biol2018; 14(11): e1006238.
[279]
TandonA, HowardB, RamaiahgariS, MaharanaA, FergusonS, ShahR, et al. Deep learning image analysis of high-throughput toxicology assay images. SLAS Discov2022; 27(1):29-38.
[280]
SchwartzMP, HouZG, PropsonNE, ZhangJ, EngstromCJ, CostaVS, et al. Human pluripotent stem cell-derived neural constructs for predicting neural toxicity. Proc Natl Acad Sci USA2015; 112(40):12516-21.
[281]
IacovacciJ, SerafiniMS, AvuzziB, BadenchiniF, CicchettiA, DevecchiA, et al. Intestinal microbiota composition is predictive of radiotherapy-induced acute gastrointestinal toxicity in prostate cancer patients. EBioMedicine2024;106:105246.
[282]
LiXN, SunY, ZhouZ, LiJR, LiuS, ChenL, et al. Deep learning-driven exploration of pyrroloquinoline quinone neuroprotective activity in Alzheimer’s disease. Adv Sci2024; 11(18):2308970.
[283]
SuR, YangHT, WeiLY, ChenSQ, ZouQ. A multi-label learning model for predicting drug-induced pathology in multi-organ based on toxicogenomics data. PLoS Comput Biol2022; 18(9):e1010402.
[284]
ZhangZX, ChenLL, WangYM, ZhangTT, ChenYC, YoonE.Label-free estimation of therapeutic efficacy on 3D cancer spheres using convolutional neural network image analysis. Anal Chem 2019; 91(21):14093-100.
[285]
WangM, MaXY, SiJW, TangHJ, WangHF, LiTL, et al. Adverse drug reaction discovery using a tumor-biomarker knowledge graph. Front Genet2021;11:625659.
[286]
AntontsevV, JagarapuA, BundeyY, HouH, KhotimchenkoM, WalshJ, et al. A hybrid modeling approach for assessing mechanistic models of small molecule partitioning in vivo using a machine learning-integrated modeling platform. Sci Rep2021; 11(1):eaax5516.
[287]
Prantil-BaunR, NovakR, DasD, SomayajiMR, PrzekwasA, IngberDE. Physiologically based pharmacokinetic and pharmacodynamic analysis enabled by microfluidically linked organs-on-chips. Ann Rev Pharmacol Toxicol2018;58:37-64.
[288]
FowlerS, ChenWLK, DuignanDB, GuptaA, HariparsadN, KennyJR, et al. Microphysiological systems for ADME-related applications: current status and recommendations for system development and characterization. Lab Chip 2020; 20(3):446-67.
[289]
PetreusT, CadoganE, HughesG, SmithA, ReddyVP, LauA, et al. Tumour-on- chip microfluidic platform for assessment of drug pharmacokinetics and treatment response. Commun Biol2021; 4(1):1001.
[290]
MaozBM, HerlandA, FitzGeraldEA, GrevesseT, VidoudezC, PachecoAR, et al. A linked organ-on-chip model of the human neurovascular unit reveals the metabolic coupling of endothelial and neuronal cells. Nat Biotechnol2018; 36 (9):865-74.
[291]
KimJ, YoonT, LeeS, KimPJ, KimY. Reconstitution of human tissue barrier function for precision and personalized medicine. Lab Chip2024; 24 (14):3347-66.
[292]
PhillipsJA, GrandhiTSP, DavisM, GautierJC, HariparsadN, KellerD, et al. A pharmaceutical industry perspective on microphysiological kidney systems for evaluation of safety for new therapies. Lab Chip2020; 20(3):468-76.
[293]
KimuraH, IkedaT, NakayamaH, SakaiY, FujiiT. An on-chip small intestine- liver model for pharmacokinetic studies. Jala2015; 20(3):265-73.
[294]
SungJH, KamC, ShulerML. A microfluidic device for a pharmacokinetic- pharmacodynamic (PK-PD) model on a chip. Lab Chip2010; 10(4):446-55.
[295]
SungJH. Pharmacokinetic-based multi-organ chip for recapitulating organ interactions. Methods Cell Biol2018;146:183-97.
[296]
HerlandA, MaozB, DasD, SomayajiMR, Prantil-BaunR, NovakR, et al. Quantitative prediction of human pharmacokinetic responses to drugs via fluidically coupled vascularized organ chips. Nat Biomed Eng2020; 4 (4):421-36.
[297]
HuhD, HamiltonGA, IngberDE.From 3D cell culture to organs-on-chips. Trends Cell Biol 2011; 21(12):745-54.
[298]
ChangSY, VoellingerJL, VanNess KP, ChapronB, ShafferRM, NeumannT, et al. Characterization of rat or human hepatocytes cultured in microphysiological systems (MPS) to identify hepatotoxicity. Toxicol In Vitro2017;40:170-83.
[299]
SkardalA, MurphySV, DevarasettyM, MeadI, KangHW, SeolYJ, et al. Multi- tissue interactions in an integrated three-tissue organ-on-a-chip platform. Sci Rep2017; 7(1):8837.
[300]
HuangZH, DentiP, MistryH, KloproggeF. Machine learning and artificial intelligence in PK-PD modeling: fad, friend, or foe? Clin Pharmacol Ther 2024; 115(4):652-4.
[301]
TranTTV, TayaraH, ChongKT. Recent studies of artificial intelligence on in silico drug distribution prediction. Int J Mol Sci2023; 24(3):1815.
[302]
TaoL, ZhangP, QinC, ChenSY, ZhangC, ChenZ, et al. Recent progresses in the exploration of machine learning methods as in-silico ADME prediction tools. Adv Drug Deliv Rev2015;86:83-100.
[303]
WuKH, LiX, ZhouZ, ZhaoYN, SuM, ChengZ, et al. Predicting pharmacodynamic effects through early drug discovery with artificial intelligence-physiologically based pharmacokinetic (AI-PBPK) modelling. Front Pharmacol2024;15:1330855.
[304]
PillaiN, DasguptaA, SudsakornS, FretlandJ, MavroudisPD. Machine learning guided early drug discovery of small molecules. Drug Discov Today2022; 27 (8):2209-15.
[305]
OtaR, YamashitaF. Application of machine learning techniques to the analysis and prediction of drug pharmacokinetics. J Control Release2022;352:352961-9.
[306]
ZouHX, BanerjeeP, LeungSSY, YanXY. Application of pharmacokinetic- pharmacodynamic modeling in drug delivery: development and challenges. Front Pharmacol2020;11:997.
[307]
ReyesDR, EschMB, EwartL, NasiriR, HerlandA, SungKY, et al. From animal testing to in vitro systems: advancing standardization in microphysiological systems. Lab Chip2024; 24(5):1076-87.
[308]
CongY, HanXH, WangYP, ChenZZ, LuY, LiuTJ, et al. Drug toxicity evaluation based on organ-on-a-chip technology: a review. Micromachines2020; 11 (4):381.
[309]
VoAH, VanVleet TR, GuptaRR, LiguoriMJ, RaoMS. An overview of machine learning and big data for drug toxicity evaluation. Chem Res Toxicol2020; 33 (1):20-37.
[310]
HartungT. Toxicology for the twenty-first century. Nature2009; 460 (7252):208-12.
[311]
MaC, ZhaoL, ZhouEM, XuJ, ShenSF, WangJY. On-chip construction of liver lobule-like microtissue and its application for adverse drug reaction assay. Anal Chem2016; 88(3):1719-27.
[312]
WuMH, HuangSB, LeeGB.Microfluidic cell culture systems for drug research. Lab Chip 2010; 10(8):939-56.
[313]
JangKJ, OtienoMA, RonxhiJ, LimHK, EwartL, KodellaKR, et al. Reproducing human and cross-species drug toxicities using a liver-chip. Sci Transl Med2019; 11(517):eaax5516.
[314]
BavliD, PrillS, EzraE, LevyG, CohenM, VinkenM, et al. Real-time monitoring of metabolic function in liver-on-chip microdevices tracks the dynamics of mitochondrial dysfunction. Proc Natl Acad Sci USA2016; 113(16):E2231-40.
[315]
XiongGL, WuZX, YiJC, FuL, YangZJ, HsiehCY, et al. ADMETlab 2.0: an integrated online platform for accurate and comprehensive predictions of ADMET properties. Nucleic Acids Res2021; 49(W1):W5-14.
[316]
LiT, TongWD, RobertsR, LiuZC, ThakkarS. DeepDILI: deep learning-powered drug-induced liver injury prediction using model-level representation. Chem Res Toxicol2021; 34(2):550-65.
[317]
RathmanJ, YangC, RibeiroJV, MostragA, ThakkarS, TongWD, et al. Development of a battery of in silico prediction tools for drug-induced liver injury from the vantage point of translational safety assessment. Chem Res Toxicol2021; 34(2):601-15.
[318]
Nguyen-VoTH, NguyenL, DoN, LePH, NguyenTN, NguyenBP, et al. Predicting drug-induced liver injury using convolutional neural network and molecular fingerprint-embedded features. ACS Omega2020; 5(39):25432-9.
[319]
KulkarniP. Prediction of drug-induced kidney injury in drug discovery. Drug Metab Rev2021; 53(2):234-44.
[320]
BaraB, GerryAL, EfthymiosK, DimitriosS, AristotelisT, VassilisGG. Machine learning approaches to predict drug efficacy and toxicity in oncology. Cells Rep Methods2023; 3(2):100413.
[321]
KujalaVJ, PasqualiniFS, GossJA, NawrothJC, ParkerKK. Laminar ventricular myocardium on a microelectrode array-based chip. J Mater Chem B2016; 4 (20):3534-43.
[322]
SinhaK, GhoshN, SilPC. A review on the recent applications of deep learning in predictive drug toxicological studies. Chem Res Toxicol2023; 36 (8):1174-205.
[323]
PetchakupC, WongSO, DalanR, HouHW. Label-free virtual staining of neutrophil extracellular traps (NETs) in microfluidics. Lab Chip2023; 23 (18):3936-44.
[324]
SchmidtBZ, LehmannM, GutbierS, NemboE, NoelS, SmirnovaL, et al. In vitro acute and developmental neurotoxicity screening: an overview of cellular platforms and high-throughput technical possibilities. Arch Toxicol2017; 91 (1):1-33.
[325]
YangJZ, RanYM, LiuSP, RenCH, LouYT, JuPF, et al. Synergistic D-amino acids based antimicrobial cocktails formulated via high-throughput screening and machine learning. Adv Sci2024; 11(9):2307173.
[326]
HuangQH, YangTH, SongYP, SunWX, XuJ, ChengY, et al. A three- dimensional (3D) liver-kidney on a chip with a biomimicking circulating system for drug safety evaluation. Lab Chip 2024; 24(6):1715-26.
[327]
BauerS, HuldtCW, KanebrattKP, DurieuxI, GunneD, AnderssonS, et al. Functional coupling of human pancreatic islets and liver spheroids on-a-chip: towards a novel human ex vivo type 2 diabetes model. Sci Rep2017; 7 (1):14620.
[328]
MontanariF, EckerGF. Prediction of drug-ABC-transporter interaction— recent advances and future challenges. Adv Drug Deliv Rev2015;86:8617-26.
[329]
NieJ, GaoQ, FuJZ, HeY.Grafting of 3D bioprinting to in vitro drug screening: a review. Adv Healthc Mater 2020; 9(7):1901773.
[330]
GötzJ, JacklMK, JindakunC, MarzialeAN, AndréJ, GoslingDJ, et al. High- throughput synthesis provides data for predicting molecular properties and reaction success. Sci Adv2023; 9(43):eadj2314.
[331]
Ortiz-PerezA, van TilborgD, van der MeelR, GrisoniF, AlbertazziL. Machine learning-guided high throughput nanoparticle design. Digit Discov2024; 3 (7):1280-91.
[332]
EgorovE, PietersC, Korach-RechtmanH, ShkloverJ, SchroederA. Robotics, microfluidics, nanotechnology and AI in the synthesis and evaluation of liposomes and polymeric drug delivery systems. Drug Deliv Transl Res2021; 11(2):345-52.
[333]
SarmadiM, BehrensAM, McHughKJ, ContrerasHTM, TochkaZL, LuXG, et al. Modeling, design, and machine learning-based framework for optimal injectability of microparticle-based drug formulations. Sci Adv2020; 6(28): eabb6594.
[334]
RebolloR, OyounF, CorvisY, El-HammadiMM, SaubameaB, AndrieuxK, et al. Microfluidic manufacturing of liposomes: development and optimization by design of experiment and machine learning. ACS Appl Mater Interfaces2022; 14(35):39736-45.
[335]
DamiatiSA, DamiatiS. Microfluidic synthesis of indomethacin-loaded PLGA microparticles optimized by machine learning. Front Mol Biosci2021;8:677547.
[336]
DamiatiSA, RossiD, JoenssonHN, DamiatiS. Artificial intelligence application for rapid fabrication of size-tunable PLGA microparticles in microfluidics. Sci Rep2020; 10(1):19517.
[337]
SmeraldoA, PonsiglioneAM, NettiPA, TorinoE. Artificial neural network modelling hydrodenticity for optimal design by microfluidics of polymer nanoparticles to apply in magnetic resonance imaging. Acta Biomater2023;171:440-50.
[338]
RoucoH, Diaz-RodriguezP, Rama-MolinosS, Remuñán-LópezC, LandinM. Delimiting the knowledge space and the design space of nanostructured lipid carriers through artificial intelligence tools. Int J Pharm2018; 553(1- 2):522-30.
SuZN, HeJX, ZhouPP, HuangL, ZhouJH. A high-throughput system combining microfluidic hydrogel droplets with deep learning for screening the antisolvent-crystallization conditions of active pharmaceutical ingredients. Lab Chip2020; 20(11):1907-16.
[341]
WuHS, HeJB, ChengHR, YangL, ParkHJ, LiJL. Development and analysis of machine-learning guided flash nanoprecipitation (FNP) for continuous chitosan nanoparticles production. Int J Biol Macromol2022;222:1229-37.
[342]
DongH, LinJW, TaoYH, JiaY, SunLN, LiWJ, et al. AI-enhanced biomedical micro/nanorobots in microfluidics. Lab Chip2024; 24(5):1419-40.
[343]
KimmigJ, SchuettT, VollrathA, ZechelS, SchubertUS.Prediction of nanoparticle sizes for arbitrary methacrylates using artificial neuronal networks. Adv Sci 2021; 8(23):2102429.
[344]
DongSH, YuHL, PoupartP, HoEA. Gaussian processes modeling for the prediction of polymeric nanoparticle formulation design to enhance encapsulation efficiency and therapeutic efficacy. Drug Deliv Transl Res2024;15:372-88.
[345]
ShengH, ChenLW, ZhaoYP, LongXA, ChenQS, WuCY, et al. Closed, one-stop intelligent and accurate particle characterization based on micro-Raman spectroscopy and digital microfluidics. Talanta2024;266:124895.
[346]
RizkinBA, ShkolnikAS, FerraroNJ, HartmanRL. Combining automated microfluidic experimentation with machine learning for efficient polymerization design. Nat Mach Intell2020; 2(4):200-9.
[347]
AandD, RanaA, MottafeghA, KimDP, SinghAK. Autonomous closed-loop photochemical reaction optimization for the synthesis of various angiotensin II receptor blocker molecules. React Chem Eng2024; 9(9):2427-35.
[348]
ColeyCW, ThomasIII DA, LummissJAM, JaworskiJN, BreenCP, SchultzV, et al. A robotic platform for flow synthesis of organic compounds informed by AI planning. Science2019; 365(6453):557.
[349]
GalanEA, ZhaoHR, WangXK, DaiQH, HuckWTS, MaSH. Intelligent microfluidics: the convergence of machine learning and microfluidics in materials science and biomedicine. Matter2020; 3(6):1893-922.
[350]
VolkAA, EppsRW, YonemotoDT, MastersBS, CastellanoFN, ReyesKG, et al. AlphaFlow: autonomous discovery and optimization of multi-step chemistry using a self-driven fluidic lab guided by reinforcement learning. Nat Commun2023;14:1403.
[351]
ZhouZP, LiXC, ZareRN. Optimizing chemical reactions with deep reinforcement learning. ACS Cent Sci2017; 3(12):1337-44.
[352]
KatesPA, CookJN, GhanR, NguyenHJ, SitasuwanP, LeeLA. Incorporation of automated buffer exchange empowers high-throughput protein and plasmid purification for downstream uses. SLAS Technol2023; 28(4):243-50.
[353]
LeeTY, ChoiTM, ShimTS, FrijnsRAM, KimSH. Microfluidic production of multiple emulsions and functional microcapsules. Lab Chip2016;163:415-40.
[354]
Vladisavljevic´GT, KhalidN, NevesMA, KuroiwaT, NakajimaM, UemuraK, et al. Industrial lab-on-a-chip: design, applications and scale-up for drug discovery and delivery. Adv Drug Deliv Rev2013; 65(11-12):1626-63.
[355]
McDonaldSM, AugustineEK, LannersQ, RudinC, BrinsonLC, BeckerML. Applied machine learning as a driver for polymeric biomaterials design. Nat Commun2023;14:4838.
[356]
VecchioG, FenechM, PompaPP, VoelckerNH. Lab-on-a-chip-based high- throughput screening of the genotoxicity of engineered nanomaterials. Small2014; 10(13):2721-34.
[357]
MahmudMM, PandeyN, WinklesJA, WoodworthGF, KimAJ. Toward the scale-up production of polymeric nanotherapeutics for cancer clinical trials. Nano Today2024;56:102314.
[358]
ShalabyKS, SolimanME, CasettariL, BonacucinaG, CespiM, PalmieriGF, et al. Determination of factors controlling the particle size and entrapment efficiency of noscapine in PEG/PLA nanoparticles using artificial neural networks. Int J Nanomed2014;949:53-64.
[359]
DengJY, YeZYF, ZhengWW, ChenJ, GaoHS, WuZ, et al. Machine learning in accelerating microsphere formulation development. Drug Deliv Transl Res2023; 13(4):966-82.
[360]
MihandoostS, RezvantalabS, PallaresRM, SchulzV, KiesslingF. A generative adversarial network approach to predict nanoparticle size in microfluidics. ACS Biomater Sci Eng2024; 11(1):268-79.
[361]
Rodríguez-DoradoR, LandínM, AltaiA, RussoP, AquinoRP,Del Gaudio P. A novel method for the production of core-shell microparticles by inverse gelation optimized with artificial intelligent tools. Int J Pharm2018; 538(1- 2):97-104.
[362]
ZaslavskyJ, BanniganP, AllenC. Re-envisioning the design of nanomedicines: harnessing automation and artificial intelligence. Expert Opin Drug Deliv2023; 20(2):241-57.
[363]
LvHL, ChenXY. Intelligent control of nanoparticle synthesis through machine learning. Nanoscale2022; 14(18):6688-708.
LiuY, YangGZ, HuiY, RanaweeraS, ZhaoCX. Microfluidic nanoparticles for drug delivery. Small2022; 18(36):2106580.
[366]
PilkingtonCP, GispertI, ChuiSY, SeddonJM, ElaniY. Engineering a nanoscale liposome-in-liposome for in situ biochemical synthesis and multi-stage release. Nat Chem2024; 16(10):1612-20.
[367]
WangHJ, LanZY, TianRZ, LiangX, JiaFH, MaM, et al. Combined helical- blade-strengthened co-flow focusing and high-throughput screening for the synthesis of highly homogeneous nanoliposomes. Nano Today2024;56:102301.
[368]
HeY, YeZYF, LiuXY, WeiZJ, QiuF, LiHF, et al. Can machine learning predict drug nanocrystals? J Control Release 2020;322:274-85.
WangDY, YangZ, ZhuBQ, MeiXF, LuoXM. Machine-learning-guided cocrystal prediction based on large data base. Cryst Growth Des2020; 20 (10):6610-21.
[371]
KimY, JungH, KumarS, PatonRS, KimS. Designing solvent systems using self-evolving solubility databases and graph neural networks. Chem Sci2024; 15(3):923-39.
[372]
MaYM, GaoZG, ShiP, ChenMY, WuSG, YangC, et al. Machine learning-based solubility prediction and methodology evaluation of active pharmaceutical ingredients in industrial crystallization. Front Chem Sci Eng2022; 16 (4):523-35.
[373]
BoobierS, HoseDRJ, BlackerAJ, NguyenBN. Machine learning with physicochemical relationships: solubility prediction in organic solvents and water. Nat Commun2020;11:5753.
[374]
GaoP, ZhangJ, SunYZ, YuJG. Accurate predictions of aqueous solubility of drug molecules via the multilevel graph convolutional network (MGCN) and SchNet architectures. Phys Chem Chem Phys2020; 22(41):23766-72.
[375]
LeeS, LeeM, GyakKW, KimSD, KimMJ, MinK. Novel solubility prediction models: molecular fingerprints and physicochemical features vs graph convolutional neural networks. ACS Omega2022; 7(14):12268-77.
[376]
VenkatramS, BatraR, ChenLH, KimC, SheltonM, RamprasadR. Predicting crystallization tendency of polymers using multifidelity information fusion and machine learning. J Phys Chem B2020; 124(28):6046-54.
[377]
BarjatH, CheckleyS, ChituT, DawsonN, FarshchiA, FerreiraA, et al. Demonstration of the feasibility of predicting the flow of pharmaceutically relevant powders from particle and bulk physical properties. J Pharm Innov2021; 16(1):181-96.
[378]
SunGX, LiuXT, AbramovYA, LillSON, ChangC, BurgerV, et al. Current state- of-the-art in-house and cloud-based applications of virtual polymorph screening of pharmaceutical compounds: a challenging case of AZD1305. Cryst Growth Des2021; 21(4):1972-83.
[379]
BaoZQ, BuftonJ, HickmanRJ, Aspuru-GuzikA, BanniganP, AllenC. Revolutionizing drug formulation development: the increasing impact of machine learning. Adv Drug Deliv Rev2023;202:115108.
[380]
BanniganP, AldeghiM, BaoZQ, HäseF, Aspuru-GuzikA, AllenC. Machine learning directed drug formulation development. Adv Drug Deliv Rev2021;175:113806.
[381]
GaoHS, KanS, YeZYF, FengYC, JinL, ZhangXD, et al. Development of in silico methodology for siRNA lipid nanoparticle formulations. Chem Eng J2022;442:136310.
[382]
DalalRJ, OviedoF, LeydenMC, ReinekeTM. Polymer design via SHAP and Bayesian machine learning optimizes pDNA and CRISPR ribonucleoprotein delivery. Chem Sci2024; 15(19):7219-28.
[383]
HanR, YangYL, LiXS, OuyangDF. Predicting oral disintegrating tablet formulations by neural network techniques. Asian J Pharm Sci2018; 13 (4):336-42.
[384]
Jones-SalkeyO, YuleC, IngramA, StahlerL, NicusanAL, CliffordS, et al. Using AI/ML to predict blending performance and process sensitivity for continuous direct compression (CDC). Int J Pharm2024;651:123796.
[385]
SharallahOA, PoddarNK, AlwadanOA.Delineation of the role of G6PD in Alzheimer’s disease and potential enhancement through microfluidic and nanoparticle approaches. Ageing Res Rev 2024;99:102394.
[386]
GaoM, LiuSB, ChenJN, GordonKC, TianF, McGoverinCM. Potential of Raman spectroscopy in facilitating pharmaceutical formulations development—an AI perspective. Int J Pharm2021;597:120334.
[387]
MunSG, ChoiHW, LeeJM, LimJH, HaJH, KangMJ, et al. rGO nanomaterial- mediated cancer targeting and photothermal therapy in a microfluidic co- culture platform. Nano Converg2020; 7(1):10.
[388]
LiangJ, LuXX, ChangTY, CuiHL. Deep learning aided quantitative analysis of anti-tuberculosis fixed-dose combinatorial formulation by terahertz spectroscopy. Spectrochim Acta A Mol Biomol Spectrosc2022;269:120746.
[389]
GalataDL, FarkasA, KönyvesZ, MészárosLA, SzabóE, CsontosI, et al. Fast, spectroscopy-based prediction of in vitro dissolution profile of extended release tablets using artificial neural networks. Pharmaceutics2019; 11 (8):400.
[390]
Henser-BrownhillT, MartinL, SamangoueiP, LadakA, ApostolidouM, NagelB, et al. In silico screening accelerates nanocarrier design for efficient mRNA delivery. Adv Sci2024; 11(30):2401935.
[391]
WangS, DiJW, WangD, DaiXD, HuaYB, GaoX, et al. State-of-the-art review of artificial neural networks to predict, characterize and optimize pharmaceutical formulation. Pharmaceutics2022; 14(1):183.
[392]
HickmanRJ, BanniganP, BaoZQ, Aspuru-GuzikA, AllenC. Self-driving laboratories: a paradigm shift in nanomedicine development. Matter2023; 6 (4):1071-81.
[393]
StrubleTJ, AlvarezJC, BrownSP, ChytilM, CisarJ, DesJarlaisRL, et al. Current and future roles of artificial intelligence in medicinal chemistry synthesis. J Med Chem2020; 63(16):8667-82.
[394]
SeglerMHS, PreussM, WallerMP. Planning chemical syntheses with deep neural networks and symbolic AI. Nature2018; 555(7698):604-10.
[395]
Sanchez-LengelingB, Aspuru-GuzikA. Inverse molecular design using machine learning: generative models for matter engineering. Science2018; 361(6400):360-5.
[396]
KlucznikT, Mikulak-KlucznikB, McCormackMP, LimaH, SzymkucS, BhowmickM, et al. Efficient syntheses of diverse, medicinally relevant targets planned by computer and executed in the laboratory. Chem2018; 4 (3):522-32.
[397]
SeglerMHS, WallerMP. Neural-symbolic machine learning for retrosynthesis and reaction prediction. Chemistry2017; 23(25):5966-71.
[398]
Strieth-KalthoffF, SzymkucS, MolgaK, Aspuru-GuzikA, GloriusF, GrzybowskiBA. Artificial intelligence for retrosynthetic planning needs both data and expert knowledge. J Am Chem Soc2024; 146(16):11005-17.
NielsenMK, AhnemanDT, RieraO, DoyleAG. Deoxyfluorination with sulfonyl fluorides: navigating reaction space with machine learning. J Am Chem Soc2018; 140(15):5004-8.
[401]
BayleyO, SavinoE, SlatteryA, NoëlT. Autonomous chemistry: navigating self-driving labs in chemical and material sciences. Matter2024; 7 (7):2382-98.
[402]
HsuehHT, ChouRT, RaiU, LiyanageW, KimYC, AppellMB, et al. Machine learning-driven multifunctional peptide engineering for sustained ocular drug delivery. Nat Commun2023;14:2509.
[403]
Pérez-LópezA, Torres-SuárezAI, Martín-SabrosoC, Aparicio-BlancoJ.An overview of in vitro 3D models of the blood-brain barrier as a tool to predict the in vivo permeability of nanomedicines. Adv Drug Deliv Rev 2023;196:114816.
[404]
KashaninejadN, MoradiE, MoghadasH. Micro/nanofluidic devices for drug delivery. Prog Mol Biol Transl Sci2022; 187(1):9-39.
[405]
SerovN, VinogradovV. Artificial intelligence to bring nanomedicine to life. Adv Drug Deliv Rev2022;184:114194.
[406]
SuY, ZhangBL, SunRW, LiuWF, ZhuQB, ZhangX, et al. PLGA-based biodegradable microspheres in drug delivery: recent advances in research and application. Drug Deliv2021; 28(1):1397-418.
[407]
GaoJJ, KarpJ, LangerR, JoshiN. The future of drug delivery. Chem Mater2023; 35(2):359-63.
[408]
MaekiM, KimuraN, SatoY, HarashimaH, TokeshiM. Advances in microfluidics for lipid nanoparticles and extracellular vesicles and applications in drug delivery systems. Adv Drug Deliv Rev2018;128:84-100.
[409]
AlaviSE, AlharthiS, AlaviSF, AlaviSZ, ZahraGE, RazaA, et al. Microfluidics for personalized drug delivery. Drug Discov Today2024; 29(4):103936.
[410]
VoraLK, GholapAD, JethaK, ThakurRRS, SolankiHK, ChavdaVP. Artificial intelligence in pharmaceutical technology and drug delivery design. Pharmaceutics2023; 15(7):1916.
[411]
HassanzadehP, AtyabiF, DinarvandR. The significance of artificial intelligence in drug delivery system design. Adv Drug Deliv Rev 2019;151- 152:169-90.
[412]
HashidaM. Role of pharmacokinetic consideration for the development of drug delivery systems: a historical overview. Adv Drug Deliv Rev2020;157:71-82.
[413]
XiaS, WangY. Preparation of solid-dosage nanomedicine via green chemistry route: advanced computational simulation of nanodrug solubility prediction using machine learning models. J Mol Liq2023;375:121319.
[414]
WangFJ, SangfuangN, McCoubreyLE, YadavV, ElbadawiM, OrluM, et al. Advancing oral delivery of biologics: machine learning predicts peptide stability in the gastrointestinal tract. Int J Pharm2023;634:122643.
[415]
YeZYF, YangYL, LiXS, CaoDS, OuyangDF. An integrated transfer learning and multitask learning approach for pharmacokinetic parameter prediction. Mol Pharm2019; 16(2):533-41.
[416]
SivasankaranS, JonnalagaddaS. Levonorgestrel loaded biodegradable microparticles for injectable contraception: preparation, characterization and modelling of drug release. Int J Pharm2022;624:121994.
[417]
HunterMR, CuiLL, PorebskiBT, PereiraS, SonziniS, OdunzeU, et al. Understanding intracellular biology to improve mRNA delivery by lipid nanoparticles. Small Methods2023; 7(9):2201695.
[418]
XuY, MaS, CuiH, ChenJ, XuS, GongF, et al. AGILE platform: a deep learning powered approach to accelerate LNP development for mRNA delivery. Nat Commun2024;15:6305.
[419]
HerrmannIK, WoodMJA, FuhrmannG. Extracellular vesicles as a next- generation drug delivery platform. Nat Nanotechnol2021; 16(7):748-59.
[420]
ElliottRO, HeM. Unlocking the power of exosomes for crossing biological barriers in drug delivery. Pharmaceutics2021; 13(1):122.
[421]
TianX, ZhuMT, TianYH, RammGA, ZhaoYL, NieGJ. A membrane vesicle- based dual vaccine against melanoma and Lewis lung carcinoma. Biomaterials2012; 33(26):6147-54.
[422]
GreenbergZF, GraimKS, HeM. Towards artificial intelligence-enabled extracellular vesicle precision drug delivery. Adv Drug Deliv Rev2023;199:114974.
[423]
RepeckaD, JauniskisV, KarpusL, RembezaE, RokaitisI, ZrimecJ, et al. Expanding functional protein sequence spaces using generative adversarial networks. Nat Mach Intell2021; 3(4):324-33.
[424]
BanniganP, BaoZQ, HickmanRJ, AldeghiM, HäseF, Aspuru-GuzikA, et al. Machine learning models to accelerate the design of polymeric long-acting injectables. Nat Commun2023;14:35.
[425]
GalataDL, KönyvesZ, NagyB, NovákM, MészárosLA, SzabóE, et al. Real-time release testing of dissolution based on surrogate models developed by machine learning algorithms using NIR spectra, compression force and particle size distribution as input data. Int J Pharm2021;597:120338.
[426]
ObeidS, MadzarevicM, KrkobabicM, IbricS. Predicting drug release from diazepam FDM printed tablets using deep learning approach: influence of process parameters and tablet surface/volume ratio. Int J Pharm2021;601:120507.
[427]
WangYW, ZouRR, ZhouYQ, ZhengY, PengC, LiuY, et al. Unraveling mechanisms of protein encapsulation and release in coacervates via molecular dynamics and machine learning. Chem Sci2024; 15(33):13442-51.
[428]
SuryavanshiP, WangJW, DuggalI, ManiruzzamanM, BanerjeeS. Four- dimensional printed construct from temperature-responsive self-folding feedstock for pharmaceutical applications with machine learning modeling. Pharmaceutics2023; 15(4):1266.
[429]
ImanparastA, AmeriAR, AttaranN, SharifiS, TamayolA, SazgarniaA, et al. A preclinical design approach for translation of biohybrid photosensitive nanoplatform for photodynamic therapy of breast cancer. J Control Release2025;378:543-58.
[430]
ApoorvaS, NguyenNT, SreejithKR. Recent developments and future perspectives of microfluidics and smart technologies in wearable devices. Lab Chip2024; 24(7):1833-66.
[431]
BagherifardS, TamayolA, MostafaluP, AkbariM, ComottoM, AnnabiN, et al. Dermal patch with integrated flexible heater for on demand drug delivery. Adv Healthc Mater2016; 5(1):175-84.
[432]
GraysonACR, ShawgoRS, LiYW, CimaMJ. Electronic MEMS for triggered delivery. Adv Drug Deliv Rev2004; 56(2):173-84.
[433]
PrausnitzMR, LangerR. Transdermal drug delivery. Nat Biotechnol2008; 26 (11):1261-8.
[434]
BiswasAA, DhondaleMR, AgrawalAK, SerranoDR, MishraB, KumarD.Advancements in microneedle fabrication techniques: artificial intelligence assisted 3D-printing technology. Drug Deliv Transl Res 2024; 14 (6):1458-79.
[435]
BagdeA, DevS, SriramLMK, SpencerSD, KalvalaA, NathaniA, et al. Biphasic burst and sustained transdermal delivery in vivo using an AI-optimized 3D- printed MN patch. Int J Pharm2023;636:122647.
[436]
CastroBM, ElbadawiM, OngJJ, PollardT, SongZ, GaisfordS, et al. Machine learning predicts 3D printing performance of over 900 drug delivery systems. J Control Release2021;337:530-45.
ChungAJ, KimD, EricksonD. Electrokinetic microfluidic devices for rapid, low power drug delivery in autonomous microsystems. Lab Chip2008; 8 (2):330-8.
[439]
SuriniS, AkiyamaH, MorishitaM, NagaiT, TakayamaK. Release phenomena of insulin from an implantable device composed of a polyion complex of chitosan and sodium hyaluronate. J Control Release2003; 90(3):291-301.
[440]
BenkoE, IlicIG, KristóK, RegdonJr G, CsókaI, Pintye-HódiK, et al. Predicting drug release rate of implantable matrices and better understanding of the underlying mechanisms through experimental design and artificial neural network-based modelling. Pharmaceutics2022; 14(2):228.
[441]
BehrensMR, RuderWC. Smart magnetic microrobots learn to swim with deep reinforcement learning. Adv Intell Syst2022; 4(10):2200023.
[442]
CaiMX, WangQQ, QiZY, JinDD, WuXY, XuTT, et al. Deep reinforcement learning framework-based flow rate rejection control of soft magnetic miniature robots. IEEE Trans Cybern2023; 53(12):7699-711.
[443]
NahonDM, MoerkensR, AydogmusH, LendemeijerB, Martinez-SilgadoA, SteinJM, et al. Standardizing designed and emergent quantitative features in microphysiological systems. Nat Biomed Eng2024; 8(8):941-62.
AI Summary 中Eng×
Note: Please be aware that the following content is generated by artificial intelligence. This website is not responsible for any consequences arising from the use of this content.
PDF
(18153KB)
