1. Introduction
Idiopathic pulmonary fibrosis (IPF) is a chronic, advancing, and life-threatening lung disorder. The disease features diffuse alveolitis, alveolar structural disruption, and excessive extracellular matrix (ECM) deposition, ultimately leading to irreversible interstitial fibrosis [
1]. Patients with IPF are confronted with a dismal clinical prognosis, with a median survival span of merely 3-5 years [
2]. Despite global aging driving an escalating incidence, current pharmaceutical therapies—pirfenidone (PFD) and nintedanib—only modestly delay disease progression and fail to halt or reverse fibrosis [
3]. This critical unmet medical challenge, combined with the lack of curative options, highlights the urgency of developing new IPF therapeutic strategies.
One of the pathological signatures of IPF is the excessive accumulation of ECM components, which is propelled by the atypically activated state of lung fibroblasts, their differentiation into myofibroblasts, and epithelial-mesenchymal transition (EMT) [
3,
4]. Collectively, these events result in the progressive scarring of lung tissue, the impairment of gas exchange, and ultimately, failure of respiratory [
2]. Although the exact mechanisms underlying IPF are not fully understood, transforming growth factor-β (TGF-β) is recognized as a central regulator of the pulmonary fibrotic processes [
2,
5-
8]. During TGF-β signaling transduction, TGF-β1 attaches to type II TGF-β receptor (TβR II), leading to the activation of TβR I (also known as ALK5 in most cell types) and phosphorylation of small mother against decapentaplegic 2 (SMAD2) and SMAD3 [
5,
9-
14]. Given TGF-β’s pivotal role in mediating EMT, fibroblast transdifferentiation, and excessive ECM accumulation, we hypothesized that targeting its downstream signaling cascade (TGF-β/ALK5) could mitigate fibrosis. Therefore, our study focused on screening inhibitors of this pathway.
Biological intelligence (BI)—the adaptive capacity of organisms to produce specialized bioactive compounds—enables sophisticated responses to environmental challenges. Medicinal plants, in particular, generate secondary metabolites with potent therapeutic potential. For example, traditional Chinese medicine formulations (e.g., Xuefu Zhuyu decoction, Maimendong decoction, and ShaShen-MaiDong decoction) exhibit anti-fibrotic efficacy, with ShaShen-MaiDong shown to suppress pulmonary fibrosis by inhibiting the TGF-β/SMAD3 pathway [
15-
18]. These findings highlight the promise of plant-derived natural products, with multi-target mechanisms and high biocompatibility, as a rich source for developing novel anti-fibrotic therapies.
Artificial intelligence (AI) and machine learning (ML) are transformative tools in the field of pharmaceuticals research and development. Over the past decades, they have revolutionized drug discovery by enabling high-throughput prediction of compound effectiveness, toxicity, and mechanism [
19]. Within the realm of IPF research, ML approaches have demonstrated notable progress in facilitating the design of novel compounds and drug repurposing opportunities [
20,
21]. Building on these advancements, we developed an ML-based classification model specifically tailored to identify inhibitors of the canonical TGF-β signaling cascade from natural products. Combined with a dual luciferase reporter assay, dihydromyricetin (DHM), a natural flavonoid derived from
Ampelopsis species, emerged as the most promising candidate from a comprehensive herbal natural product library. Subsequent experiments revealed DHM suppressed TGF-β1-induced EMT in A549 cells, fibroblast transdifferentiation in medical research council cell strain 5 (MRC-5) cells, and bleomycin (BLM)-induced pulmonary fibrosis in mice. Mechanistic studies showed DHM directly interacted with ALK5, inhibited its kinase activity, and reduced cell membrane ALK5 expression, thereby disrupting the TGF-β/ALK5 signaling axis.
Collectively, these findings suggest, for the first time, that DHM alleviates pulmonary fibrosis in mice and suppresses EMT and fibroblast transdifferentiation by repressing the TGF-β/ALK5 signaling pathway, potentially through competitive kinase inhibition and ALK5 downregulation. Overall, DHM, identified via ML-based screening and validated in luciferase reporter assays, appears to be a promising therapeutic candidate for IPF treatment. This study highlights the potential of integrating computational and experimental approaches to discover novel therapies for complex lung disorders.
2. Materials and methods
2.1. Reagents
In-house natural product group was obtained from Herbpurify (China). Plasmid pGL4.48[luc2P/SBE/Hygro], plasmid pGL4.74[hRluc/ TK], FuGENE HD transfection reagent, and Dual-Glo Luciferase Assay System were obtained from Promega (USA). Opti-minimal essential medium (MEM) was obtained from Thermo Fisher (USA). DHM was purchased from Herbal Source Biotechnology (China). BLM was provided by Nippon Kayaku (Japan). PFD was provided by Meilunbio (China). Recombinant human TGF-β1 was obtained from Peprotech (USA). Antibodies against fibronectin, α-smooth muscle actin (α-SMA), SMAD2, SMAD3, collagen type I α 1 (COL1A1), phosphorylated SMAD2 (p-SMAD2), p-SMAD3, TGF-β1, and TβR I (ALK5) were obtained from Abcam (USA). Primary antibodies against N- cadherin, matrix metalloproteinase-2 (MMP-2), protein kinase B (Akt), extracellular signal-regulated protein kinase 1/2 (Erk1/2), stress activated protein kinase (SAPK)/c-Jun amino terminal kinase (JNK), p38 mitogen-activated protein kinase (MAPK), p-Akt, p-Erk1/2, p-SAPK/JNK, and p-p38 MAPK were purchased from Cell Signaling Technology (USA). Antibody against β-actin was purchased from Sigma-Aldrich (USA). Antibodies against mice and rabbits were obtained from ZSGB-BIO (China). Recombinant human TβR I (ALK5) protein (Glu162-Ala403) was obtained from Kangbei Technology (China). SB431542 was obtained from Sigma-Aldrich.
2.2. ML-based classification model architecture
The training data for our model included parameters representing the affinity between pathway proteins and ligands. Because of their important functions, TGF-β1, SMAD2, TβR I, SMAD3, and TβR II were chosen as pivotal target proteins for researching the TGF-β signaling pathway [
9,
10,
22,
23]. Inhibitors were collected from BindingDB [
24]. All compounds were converted into canonical SMILES using RDKit [
25]. Assay types, including inhibition constant (
Ki), dissociation constant (
Kd), half maximal inhibitory concentration (IC
50), and half effective concentration (EC
50), were incorporated into the training dataset, with a threshold of 10 000 nmol·L
-1 set to define positive samples, resulting in affinity data for a total of 6642 relevant compounds. The same process was applied to the ChEMBL database, which was deduplicated against the BindingDB, and the resulting data was stored as an external test set [
26]. Therefore, we ensure that our model evaluation is based on distinct and non-redundant data, allowing for a more accurate assessment of the model’s predictive capabilities for TGF-β pathway inhibitors.
In our model construction, the Uni-Mol model, which utilizes a Transformer-based architecture equipped with self-attention mechanisms to analyze structured data effectively and establish goal-oriented relationships [
27], was employed as the pretrained foundation. Specifically, our model was fine-tuned by the standard pretrained Uni-Mol model with the checkpoint file mol_pre_no_h_220816.pt, ensuring that our fine-tuning process built upon this extensive prior knowledge to enhance the prediction accuracy of pathway inhibitors. Starting with a pretrained model, the output layer of the target model was configured to match the number of categories in the target dataset. During training, the output layer was trained from scratch, while the parameters of all the other layers were fine-tuned based on the parameters of the source model. The model was trained with a batch size of 16 over 40 epochs, optimizing both computational efficiency and comprehensive learning. A learning rate of 0.0001 was set, combined with a warm-up ratio of 0.03.
In our model evaluation, the BindingDB data was utilized for both the training and validation of the Uni-Mol model to ensure that it learns meaningful representations relevant to molecular interactions. To evaluate the robustness and stability of the model, a five-fold cross-validation approach was employed. In each iteration of the cross-validation, four of these parts were utilized for model training, while the remaining segment was reserved for validation. This procedure was repeated five times, allowing every part of the dataset to serve as the validation set exactly once. The primary evaluation metric for assessing the model's performance was the area under the precision-recall curve (AUPRC). After the completion of the cross-validation, the conclusive assessment of the model’s performance was conducted using the entire dataset. In addition to the internal cross-validation framework, the study also introduced ChEMBL as an external test set.
2.3. Dual-luciferase reporter assay
This assay was conducted as previously described [
28,
29].
2.4. Cell lines and cell culture
The human alveolar basal epithelial cell line A549, the human lung fibroblast cell line MRC-5, the kidney epithelial cell line Vero, and the human liver cancer cell line HepG2 were obtained from the American Type Culture Collection (USA). A549 and Vero cells were cultured in Dulbecco’s modified Eagle’s medium (DMEM; Gibco, USA) with 10% fetal bovine serum (FBS; Gibco), plus 1% penicillin and streptomycin (P/S; Gibco); MRC-5 and HepG2 cells were cultured in MEM (Gibco) with 10% FBS, plus 1% non-essential amino acid (NEAA; Gibco) and 1% P/S. All cells were cultivated in a moist incubator set at 37 °C with a 5% CO2 atmosphere (Thermo Fisher).
2.5. Cell-related functional assays
Cell viability assay, scratch wound healing assay, transwell migration assay, and western blotting assay were performed in accordance with the methodology detailed in our previous study [
30].
2.6. Animal experiment procedures
All animal procedures were conducted in compliance with ethical guidelines and were approved by the Laboratory Animal Ethics Committee of the Institute of Medicinal Biotechnology, Chinese Academy of Medical Sciences (approval No. IMB-20211123D1). Throughout the study, we adhered to the principles of the replacement, reduction, and refinement (3Rs) to ensure animal welfare. Male C57BL/6J mice (16-18 g) were provided by SPF (Beijing) Biotechnology Co., Ltd. (certification No. SCXK-2019-0010; China) and were housed in specific pathogen-free (SPF) facilities with unrestricted access to fresh food and water. To induce pulmonary fibrosis, on day 0, mice were put under anesthesia with tribromoethanol (250 mg·kg-1, intraperitoneally (i.p.); Sigma-Aldrich). Subsequently, BLM (5 mg·kg-1) was instilled intratracheally to the anesthetized mice. On day 7 post-BLM administration, animals were randomly allocated into four experimental groups (n = 22 per group): ① BLM model group (vehicle control); ② PFD treatment group (50 mg·kg-1); ③ low-dose DHM group (25 mg·kg-1); and ④ high-dose DHM group (50 mg·kg-1). From days 7 to 28, treatment groups received daily oral gavage of either DHM (at respective doses) or PFD, in contrast, control animals were administered an equal volume of 0.5% sodium carboxymethylcellulose vehicle. Body weights and clinical signs were monitored throughout the experimental period.
2.7. Histological assessment
The left lung of mice was prepared for histological assessment as Li et al.’s study [
31]. The pathological results were evaluated by pathologists to evaluate the degree of pulmonary fibrosis according to Ashcroft’s fibrosis score [
32].
2.8. Respiratory functions
After being anesthetized and intratracheally intubated, the mice were connected to the FlexiVent Pulmonary Respiratory Function Analysis System (SCIREQ, France) to assess the indicators of respiratory system compliance, lung respiratory volume, quasi-static compliance, respiratory system resistance, tissue respiration resistance, and respiration system elasticity.
2.9. Inflammatory cell analysis
Bronchoalveolar lavage fluid (BALF) were centrifuged at 3500 r·min-1 for 10 min at 4 °C to pellet cellular components. The cell pellets were gently resuspended in 100 μL of sterile saline. Total inflammatory cell counts were quantified by means of an automated hematology analyzer (ABX Pentra 60; HORIBA Medical, France) according to standard protocols.
2.10. Hydroxyproline assay
Hydroxyproline assay was performed as described by Li et al. [
31].
2.11. Western blotting-cellular thermal shift assay (WB-CETSA)
This assay was performed according to the methodology described by Li et al. [
31].
2.12. Molecular docking stimulation
The potential binding modes between DHM and the ALK5 kinase domain were investigated through computational docking simulations. Using AutoDock Vina 1.2.2 [
33], we docked the three-dimensional (3D) structure of DHM (obtained from PubChem) into the binding site of the human ALK5 kinase domain (retrieved from RCSB Protein Data Bank (PDB), accession code 6B8Y, resolution 0.375 Å). The resulting docking poses were evaluated and visualized using Discovery Studio 2019 to identify the most probable binding conformations.
2.13. Molecular dynamics simulations
To assess the stability of the DHM-ALK5 complex, we conducted all-atom molecular dynamics simulations with the following parameters: ① System preparation: Applied the AMBER99SB-ILDN force field for protein atoms [
34]; generated DHM parameters using general amber force field (GAFF) through Sobtop [
35]; implemented transferable intermolecular potential 3 points (TIP3Ps) explicit water model for solvation; created a cubic simulation box with 0.8 nm minimum edge distance [
36,
37]; and neutralized the system with appropriate Na
+/Cl
-. ② Simulation protocol: Performed energy minimization (steepest descent, 5000 steps); conducted constrained equilibration (500 ps) with gradual temperature increase to 298.15 K; executed production runs using the Verlet integration scheme (2 fs timestep) [
38]; maintained at 298.15 K and 1 bar (1 bar = 100 kPa) using V-rescale thermostat and Parrinello-Rahman methods [
39]; and applied periodic boundary conditions throughout the simulation. ③ Trajectory analysis: Evaluated system stability through root mean square deviation (RMSD) calculations; visualized molecular trajectories using VMD 1.9.4 [
40]; and estimated binding free energies via molecular mechanics/Poisson-Boltzmann surface area (MM/PBSA) calculations using gmx_MM/PBSA [
41,
42].
2.14. Micro-scale thermophoresis (MST) assay
The intracellular domain of recombinant human ALK5 (residues Glu162-Ala403, UniProt ID P36897) was diluted in phosphate-buffered saline (PBS; pH 7.4) to a final concentration of 5 µmol·L-1. Binding interactions were assessed using a Monolith NT.LabelFree system (NanoTemper Technologies, Germany) with premium-coated capillaries, following the manufacturer’s recommended protocol in MO.Control software. SB431542, a known TGF-β pathway inhibitor, served as the positive control. Prior to the assay, baseline fluorescence levels of the protein, test compounds, and buffer were measured to account for background signal. For binding measurements, ALK5 was maintained at 5 nmol·L-1, while excitation power was set to 5% and MST power to medium. Dose-response curves were generated to determine binding affinities, and Kd were calculated using the instrument’s built-in affinity analysis module.
2.15. ALK5 kinase enzyme assay
The inhibitory effect of DHM on ALK5 (TβR I) kinase activity was evaluated using a commercial kinase assay kit (Promega) following the supplier’s protocol. Reactions were carried out in 384-well plates, with each well containing varying concentrations of DHM, SB431542 (a reference inhibitor), or a 5% dimethyl sulfoxide (DMSO) vehicle control, along with the necessary assay components. After incubating for 30 min at room temperature, the ADP-Glo™ reagent was introduced to halt the kinase reaction and convert residual ADP to ATP. Following a 40-min incubation, the kinase detection reagent (Promega) was added to generate a luminescent signal proportional to ATP levels. The plate was then incubated for an additional 30 min before luminescence was quantified using a microplate reader. Relative kinase activity was determined based on the measured signal intensity.
2.16. Site-directed mutagenesis and transfection of ALK5 variants
Plasmids overexpressing wild-type (WT) ALK5 or its Lys232Ala, and His283Ala mutants were constructed in the pcDNA3.1(+) vector (Genscript Biotech, China). For transfection, A549 cells were seeded in six-well plates or 10-cm dishes and cultured in DMEM supplemented with 10% FBS. After 24 h, the cells were transfected with the following plasmids using polyethyleneimine (PEI), according to the manufacturer’s instructions: pcDNA3.1(+) (empty vector control), pcDNA3.1(+)-ALK5-WT-His (WT ALK5), pcDNA3.1(+)-ALK5-Lys232Ala-His (Lys232Ala mutant), and pcDNA3.1(+)-ALK5- His283Ala-His (His283Ala mutant). Following 48 h of incubation, the transfected cells were harvested for western blotting and WB-CETSA analyses.
2.17. Statistics
All numerical data are presented as mean ± standard deviation (SD), derived from a minimum of three replicate experiments. For ordinal categorical variables (e.g., pathological scores), results are reported as median with interquartile range. Data analysis was conducted by means of GraphPad Prism 9.0 and SPSS 26.0. For normally distributed data with homogeneous variance across multiple groups, one-way analysis of variance (ANOVA) was applied, followed by the student-newman-keuls (SNK) post hoc test for intergroup comparisons. In cases of heteroscedasticity, nonparametric Rank Cases transformation was performed prior to analysis. If significant differences were detected, the least significant difference (LSD) method was employed for pairwise comparisons. Survival analysis was performed using the Kaplan-Meier method, with group differences assessed via the log-rank (Mantel-Cox) test. Additionally, survival rates at defined endpoints were compared using the χ2 test or Fisher’s exact test, as appropriate. Nonparametric statistical methods were applied for ordinal categorical data. The Kruskal-Wallis H test was used for multi-group comparisons, while the Mann-Whitney U test analyzed differences between two independent samples. A p-value < 0.05 was deemed statistically significant.
3. Results
3.1. An ML-based classification model was established for TGF-β/SMAD signaling pathway inhibitors prediction
The present study developed a screening model for identifying inhibitors of the TGF-β/SMAD signaling cascade by fine-tuning the pretrained model Uni-Mol. Compounds targeting TGF-β1, SMAD2, TβR II, SMAD3, and TβR I were imported, including a total of 3226 compounds with 6642 sets of compound activity data from the BindingDB database, and 1623 compounds with 2777 sets of compound activity data from the ChEMBL database (
Fig. 1(a)).
BindingDB and ChEMBL datasets were utilized to assess the model’s performance. The ChEMBL dataset was used for external validation with molecules different from those in BindingDB. Our final model demonstrated an outstanding performance, achieving an AUPRC of 0.936 and an area under the receiver operating characteristic curve (AUROC) of 0.902, highlighting its strong discriminative power (
Table 1). The optimal classification threshold was identified as 0.2648, resulting in an accuracy of 0.787. To discover TGF-β signaling pathway inhibitors and to assess the generalization ability of our model, this screening criterion was applied to filter an herbal medicine-derived natural product library database containing 16 708 compounds presented by TargetMol, and 408 positive candidates were predicted to be potential TGF-β signaling pathway inhibitors. Through analyzing the predicted score, physicochemical properties, reported biological activities, and availability of these positive candidates, 20 compounds were selected for further identification (
Fig. 1(b), Table S1 in Appendix A).
3.2. DHM was discovered as a TGF-β/SMAD signaling pathway inhibitor
Subsequently, the inhibitory effects of these 20 compounds were confirmed by a dual-luciferase reporter assay, which is based on a plasmid containing SMAD binding element that promotes luciferase reporter gene
luc2P transcription. The feasibility of this method for detecting the effects of compounds on the TGF-β/SMAD signaling pathway is demonstrated in Fig. S1 in Appendix A. Eventually, nine compounds showed the inhibitory effects on luciferase activities (
Fig. 1(c), yellow column represents DHM and green columns represent other eight compounds), among which a hit compound, DHM (
Fig. 2(a)), exerted the most potent inhibitory effect on TGF-β/SMAD signaling cascade (Table S2 in Appendix A). Therefore, DHM was further investigated for therapeutic efficacy on IPF.
3.3. DHM inhibited TGF-β1-induced cell migration
Based on the low cytotoxic effect of DHM on various cell lines (
Fig. 2(b), Fig. S2 in Appendix A), A549 cells were pretreated with 10 to 40 μmol·L
-1 of DHM. After being induced by TGF-β1, the scratch widths of A549 were markedly decreased, while 10 to 40 μmol·L
-1 of DHM pretreatment significantly inhibited the TGF-β1 promoted cell migration at 24 and 48 h in a way that depends on concentration (
Fig. 2(c),
p < 0.05). Furthermore, a comparable outcome was noted in the transwell assay (
Fig. 2(d)). In summary, DHM inhibited A549 cell migration triggered by TGF-β1, indicating its potential inhibitory effects on EMT.
3.4. DHM repressed TGF-β1-induced EMT and fibroblast transdifferentiation by down-regulating TGF-β signaling cascade via SMAD and non-canonical pathways
TGF-β1 induces epithelial cells to acquire mesenchymal phenotypes, with cells changing from a cobblestone-like to spindle-shaped morphology. As shown in Fig. S3 in Appendix A, DHM treatment restored the morphology of TGF-β1-induced A549 cells from spindle-shaped to cobblestone-like. The expression of mesenchymal markers, which promote cells to exhibit invasive and motile capabilities, such as fibronectin,
N-cadherin, MMPs, and α-SMA, were elevated following induction by TGF-β1 (
p < 0.001). Pretreatment with DHM substantially reduced the levels of these proteins (
p < 0.05) in A549 cells (
Fig. 3(a)). In addition, the expression of COL1A1, which is an important ECM component and a fibrogenic marker correlated with the pathologic process of pulmonary fibrosis, was also significantly promoted by TGF-β1 induction in MRC-5 cells (
p < 0.0001). Similarly, the TGF-β1-enhanced expression of α-SMA, COL1A1,
N-cadherin, MMP-2, and fibronectin was significantly reduced by DHM in a concentration-dependent manner in MRC-5 cells (
Fig. 3(b),
p < 0.05).
The phosphorylation of mediators of the TGF-β/SMAD signaling cascade and TGF-β signaling cascade via non-canonical pathways were also determined by western blotting assay. Neither the pretreatment of DHM nor the induction by TGF-β1 influenced the expressions of total SMAD2, SMAD3, Erk1/2, Akt, SAPK/JNK, or p38 MAPK in A549 and MRC-5 cells (
Figs. 4(a) and
(b)). However, the ratios between phosphorylated proteins and total proteins (both phosphorylated and non-phosphorylated proteins) of these signaling proteins were significantly upregulated by TGF-β1 (
p < 0.05), which were effectively reversed by 40 μmol·L
-1 DHM pretreatment (
p < 0.05), confirming that DHM effectively repressed EMT and fibroblast transdifferentiation by down-regulating the TGF-β signaling cascade via SMAD and non-canonical pathways (
Figs. 4(c) and
(d)).
3.5. DHM ameliorated BLM-induced pulmonary fibrosis in mice by down-regulating TGF-β signaling cascade via SMAD and non-canonical pathways
To explore the potential effects of DHM on IPF, male C57BL/6J mice were intratracheally nebulized with BLM (5 mg·kg
-1) to establish an animal model of pulmonary fibrosis. The mice with obvious body weight loss were randomly grouped and treated with 50 mg·kg
-1 PFD, 25 mg·kg
-1 DHM, 50 mg·kg
-1 DHM, or an equal volume of vehicle by gavage once a day for 21 d (
Fig. 5(a)). As illustrated in
Fig. 5(b), the body weight of BLM-modeled mice was significantly reduced (
p < 0.0001). DHM administration significantly increased the body weight of BLM-induced pulmonary fibrosis mice (
p < 0.01), especially, the body weight of mice in the 25 mg·kg
-1 DHM treated group ((23.25 ± 1.18) g) recovered to the level of the control mice ((24.81 ± 0.98) g). No significant difference was observed between the PFD treated group ((21.56 ± 1.93) g) and the DHM treated group ((22.19 ± 1.52) g in the 50 mg·kg
-1 DHM group). Additionally, BLM induction led to the death of mice with a mortality of 22.7% (
p = 0.0188). The mortality of BLM-modeled pulmonary fibrosis mice was reduced by DHM administration (50 mg·kg
-1) from 22.7% to 0 (
p = 0.0188), while administration PFD (50 mg·kg
-1), did not affect the mortality (
Fig. 5(c)).
Histologically, administration of DHM effectively alleviated BLM-induced inflammation and collagen deposition in lung tissue. This was clearly demonstrated by hematoxylin and eosin (HE) staining, which showed reduced inflammatory cell infiltration, and Masson’s staining, indicating a decrease in collagen fiber accumulation (
Fig. 5(d)). Based on the results of Masson’s staining, the ratio between collagen area and tissue area was calculated. DHM and PFD administration significantly decreased the collagen area in lung tissues of mice (
Fig. 5(e),
p < 0.001). No significant difference was observed between the PFD treated group ((1.93 ± 0.75)%) and the DHM treated group ((2.94 ± 2.86)% in the 25 mg·kg
-1 DHM group, (1.38 ± 0.13)% in the 50 mg·kg
-1 DHM group). Moreover, DHM and PFD administration significantly recovered Ashcroft’s fibrosis score median of mice to the same level as the animals in the control group (
Fig. 5(f),
p = 0.0131). Additionally, the hydroxyproline assay showed that both PFD and DHM administration could significantly reduce the hydroxyproline contents in the pulmonary tissues of fibrotic mice (
Fig. 5(g),
p < 0.0001), which is consistent with the results of collagen area calculating. DHM exerted non-superior effect to PFD on reducing the hydroxyproline contents ((0.55 ± 0.15) μg·mg
-1 in the 50 mg·kg
-1 PFD group, (0.70 ± 0.09) μg·mg
-1 in the 50 mg·kg
-1 DHM group;
p = 0.0031). During pulmonary fibrosis development, mice exhibited a substantially higher lung/body weight ratio (Fig. S4(a) in Appendix A,
p < 0.0001), and the administration of both DHM and PFD significantly decreased the lung weight to body weight ratio of mice with pulmonary fibrosis to the same level as the control mice. No significant difference was observed between the PFD treated group (0.010 ± 0.002) and the DHM treated group (0.011 ± 0.002 in the 25 mg·kg
-1 DHM group, 0.010 ± 0.002 in the 50 mg·kg
-1 DHM group).
The respiratory functions of mice were assessed to reflect the therapeutic effectiveness of DHM. As shown in
Fig. 5(h), BLM treatment damaged the respiratory functions (
p < 0.05). On the one hand, the values of inspiratory capacity (IC), respiratory system compliance (
Crs), and quasi-static compliance (
K) of mice were significantly decreased by BLM induction and were recovered by 50 mg·kg
-1 of PFD and 25 mg·kg
-1 of DHM administration (
p < 0.05). On the other hand, the values of tissue respiration resistance (
G), respiratory system elasticity (
Ers), and respiratory system resistance (
Rrs) were increased substantially (
p < 0.01) in the BLM group. They were reduced in both 50 mg·kg
-1 PFD group and 50 mg·kg
-1 DHM group (
p < 0.01). Notably, IC in the 50 mg·kg
-1 DHM group ((0.74 ± 0.05) mL) was significantly higher than that in the 50 mg·kg
-1 PFD group ((0.68 ± 0.05) mL,
p = 0.0126). The differences in
Crs ((0.034 ± 0.004) vs (0.035 ± 0.007) mL·cmH
2O
-1),
K ((0.115 ± 0.005) vs (0.118 ± 0.003) cmH
2O
-1),
G ((5.82 ± 0.36) vs (5.26 ± 0.70) cmH
2O·mL
-1),
Ers ((30.12 ± 3.33) vs (32.40 ± 2.50) cmH
2O·mL
-1), and
Rrs ((0.97 ± 0.11) vs (0.94 ± 0.08) cmH
2O·s·mL
-1) between the administration of 50 mg·kg
-1 of PFD and the administration of the same dose of DHM were not significant.
Considering the important role of inflammatory reactions in pulmonary fibrosis pathologic development, inflammatory cells in the BALF of mice were quantified. As shown in Figs. S4(b)-(g) in Appendix A, the number of white blood cell (WBC), lymphocyte (LYM), monocyte (MON), neutrophil (NEU), eosinophil (EOS), and basophil (BAS) were significantly increased in lung tissues of mice because of BLM induction (p < 0.001). DHM administration at the dose of 50 mg·kg-1 decreased all inflammation-associated cells we detected (p < 0.01), while 50 mg·kg-1 of PFD showed no effects on reducing MON and BAS (p > 0.05 vs the BLM group). The number of MON ((0.12 ± 0.05) × 106 vs (0.03 ± 0.03) × 106 mL-1, p = 0.0015) and BAS ((0.013 ± 0.005) × 106 vs (0.004 ± 0.005) × 106 mL-1, p = 0.0086) was significantly higher in the 50 mg·kg-1 PFD group than the 50 mg·kg-1 DHM group.
To elucidate the potential therapeutic effectiveness of DHM administration on experimental pulmonary fibrosis and to reveal the mechanism underlying, fibronectin expression and the phosphorylation of several mediators was detected. As shown in
Figs. 5(i) and
(j), fibronectin, the fibrosis-related protein, was significantly promoted in the lung tissues of mice exposed to BLM (
p < 0.0001). Administration of PFD and DHM markedly reduced its expression (
p < 0.0001). Consistently, the phosphorylation of SMAD2, Erk1/2, SMAD3, and p38 MAPK were increased in the BLM group (
p < 0.0001), indicating that BLM induction led to the development of pulmonary fibrosis as a result of the activated TGF-β signaling cascade via SMAD and non-canonical pathways in mice. Both 25 mg·kg
-1 of DHM and 50 mg·kg
-1 of DHM administration effectively reduced the phosphorylation levels of the proteins mediating TGF-β signaling cascade via SMAD and non-canonical pathways (
p < 0.05), while 50 mg·kg
-1 of PFD administration showed no effects on the phosphorylation of SMAD3. Notably, based on the statistical analysis, DHM administration was shown to exert better efficacy in suppressing the SMAD-dependent TGF-β signaling pathway than PFD (
Fig. 5(k),
p < 0.05). The ratios of p-SMAD2/SMAD2 ((1.52 ± 0.29) vs (0.84 ± 0.50)) and p-SMAD3/SMAD3 ((2.05 ± 0.25) vs (1.69 ± 0.36)) were significantly higher in the 50 mg·kg
-1 PFD than the 50 mg·kg
-1 DHM group. DHM administration significantly reduced the ratios of p-Erk1/2/Erk1/2 and p-p38 MAPK/p38 MAPK (
Fig. 5(k),
p < 0.0001). No significant differences were observed in the inhibitory effects between DHM and PFD treatment in the phosphorylation of Erk1/2 and p38 MAPK.
In the in vivo safety assessment, no abnormal changes in body weight, survival, or obvious impairment of liver and kidney functions were observed in mice treated with a high dose of DHM (100-1000 mg·kg-1). In contrast, PFD (100-1000 mg·kg-1) treatment slowed down the body weight gain of mice, with this effect being most pronounced at 24 h post-administration (Figs. S5(a) and (b) in Appendix A, p < 0.05); among these groups, the survival rate of mice in the 1 g·kg-1 PFD group decreased to 60% (Figs. S5(c) and (d) in Appendix A, p = 0.0033 vs the 1 g·kg-1 DHM group). At 24 h post-administration, there were no significant differences in the serum levels of alanine aminotransferase (ALT), aspartate aminotransferase (AST), creatinine (Cre), and urea nitrogen (Urea) among mice in different groups (Figs. S5(e)-(h) in Appendix A). However, at 7 d post-administration, the serum ALT and AST levels in mice treated with 0.5-1 g·kg-1 PFD were significantly higher than those in the DHM treatment group (Figs. S5(i) and (j) in Appendix A, p < 0.0001). No significant differences were observed in the serum Cre and Urea among mice in different groups at 7 d post-administration (Figs. S5(k) and (l) in Appendix A).
3.6. DHM down-regulated TGF-β signaling cascade via ALK5
WB- CETSA quantifies variations in the thermal stability of the target protein when ligand binding occurs, therefore, reflecting the binding between ligand and protein. As shown in
Fig. 6(a), ALK5 was more stable in the presence of DHM, suggesting that ALK5 protein was the binding target of DHM. Molecular docking analysis was performed to predict the possible interaction (predicted binding energy = -9.60 kcal·mol
-1) between DHM and the catalytic domain of ALK5 (PDB ID: 6B8Y). Molecular dynamics simulations were then used to further evaluate the binding stability and interaction energies. Based on MM/PBSA calculations, the binding free energy of the interaction between DHM and ALK5 is (-35.20 ± 0.17) kcal·mol
-1. DHM forms hydrogen bonds with His283 and Lys232, π-alkyl with Leu340, Ile211, Ala230, and π-lone pair with Ser280 (
Figs. 6(b) and
(c)). Subsequently, MST assay confirmed the direct binding affinity between DHM and the intracellular region of ALK5 with a
Kd value of (5.09 ± 0.20) μmol·L
-1 (
Fig. 6(d)). Finally, DHM demonstrated dose-dependent inhibitory effects on ALK5 kinase activity across varying ATP concentrations (2.5, 5.0, and 10.0 μmol·L
-1), the physiological substrate of this kinase. The calculated IC
50 values exhibited a significant ATP concentration-dependent increase ((34.00 ± 4.91) μmol·L
-1 at 2.5 μmol·L
-1 ATP, (53.31 ± 10.43) μmol·L
-1 at 5 μmol·L
-1 ATP, and (112.33 ± 2.05) μmol·L
-1 at 10 μmol·L
-1 ATP;
Fig. 6(e)). This characteristic positive correlation between ATP concentration and IC
50 magnitude establishes DHM as a competitive inhibitor of ALK5 kinase (Fig. S6 in Appendix A). In addition, the pretreatment of DHM showed inhibitory effects on the expression of ALK5 protein on the cell membrane (
Fig. 6(f),
p < 0.05). Site-directed mutagenesis of Lys232 and His283 in ALK5 was accomplished using exogenous plasmids (
Fig. 6(g)). WB-CETSA revealed that DHM binds to WT
ALK5, whereas mutations at either Lys232Ala or His283Ala abolished this interaction (
Fig. 6(h)). Overexpression of ALK5 reversed the inhibitory effect of DHM on TGF-β1-induced EMT, as shown by the increased expression of the biomarkers fibronectin and
N-cadherin (
Fig. 6(i)).
4. Discussion
Although the exact mechanism of IPF remains incompletely understood, substantial evidence implicates TGF-β1 as a key player in its pathogenesis. Studies have demonstrated elevated
TGF-β1 messenger RNA (mRNA) expression in IPF patients’ alveolar macrophages and increased TGF-β1 levels in BLM-induced animal models [
43-
47]. The TGF-β signaling pathway is substantially upregulated in IPF, driving fibrogenic responses through alveolar epithelial cell injury, EMT, fibrocyte recruitment, and myofibroblast differentiation [
8]. These processes promote the transcription of fibrosis-related genes, leading to excessive ECM deposition, including collagens, proteoglycans, and glycoproteins (such as fibronectin and laminin), which compromise lung tissue integrity and function [
47-
49]. Beyond TGF-β, various signaling pathways including Wnt/β-catenin, Notch, nuclear factor κB (NF-κB), prostaglandin E2 (PGE2), mammalian target of rapamycin (mTOR), lysophosphatidic acid (LPA), and vascular endothelial growth factor receptor (VEGFR) that regulate development, senescence, inflammation, immune response, oxidation, apoptosis, autophagy, and angiogenesis, also contribute to IPF [
50-
53]. Nevertheless, the TGF-β pathway remains central to the fibrotic regulatory network, making it a prime target for therapeutic intervention.
Given the complexity of IPF pathogenesis and the pivotal role of the TGF-β1 pathway, there remains an urgent need for innovative therapeutic strategies. Natural products derived from Chinese herbal medicines represent a promising resource, offering a diverse array of bioactive compounds—including flavonoids, alkaloids, terpenoids, and polysaccharides—that may modulate the TGF-β1-driven fibrotic cascade, either directly or indirectly. From an evolutionary perspective, many of these compounds have developed specific biochemical functions, such as flavonoids, which protect plants from oxidative stress [
54]. Intriguingly, this inherent antioxidant capacity may also counteract oxidative damage in pulmonary fibrosis, highlighting their therapeutic potential [
55].
Traditional drug discovery faces significant challenges, including high costs (about 2.6 billion USD per drug), prolonged development timelines (12-15 years), and low clinical success rates (< 10%) [
56]. Under these circumstances, AI-assisted drug screening has emerged as a paradigm-shifting approach, capable of analyzing complex biological data, predicting ligand-receptor interactions, optimizing molecular design, and forecasting pharmacokinetic properties [
57]. AI not only accelerates drug discovery but also enhances efficiency by narrowing candidate selection and enabling drug repurposing. A notable example is the AI-derived compound INS018_055, developed by Insilico Medicine for IPF treatment, which has demonstrated favorable safety in Phase I trials (NCT05154240) and is now under Phase II evaluation (NCT05938920) [
58-
60]. This breakthrough underscores AI’s potential in modern drug development and aligns with our investigative approach.
In this investigation, we employed a classification model based on ML, trained on known TGF-β signaling pathway inhibitors, to screen a herbal medicine-derived natural product library database. From 408 predicted compounds, 20 were selected for experimental validation using a dual-luciferase reporter system responsive to TGF-β1 stimulation. Nine compounds showed the ability to inhibit the TGF-β/SMAD pathway, demonstrating the model’s generalization capability (45% accuracy). Notably, DHM, a polyhydroxy flavonoid from Ampelopsis grossedentata, emerged as the most potent inhibitor.
Historically,
Ampelopsis grossedentata has been utilized in traditional medicine for its diverse therapeutic effects, including hepatoprotection, anti-inflammatory activity, and metabolic regulation [
61-
63]. DHM, its principal bioactive constituent, has long been used for various ailments and exhibits diverse pharmacological effects, including reducing oxidative and inflammatory reactions and organ-protective effects against tumor [
64-
71]. Recent studies have highlighted its potential in mitigating renal and liver fibrosis [
72-
75], with preliminary evidence suggesting efficacy in pulmonary fibrosis. Xiao et al. [
76] preliminarily reported the efficacy of DHM in alleviating experimental pulmonary fibrosis in mice. Li et al. [
77] noted that higher concentrations of DHM (100 to 300 μmol·L
-1) could alleviate pulmonary fibrosis by modulating signal transducer and activator of transcription 3 (STAT3)/p-STAT3/glucose transporter 1 (GLUT1) signaling pathway
in vitro . However, its effect on TGF-β signaling remained unexplored.
Our findings reveal that DHM downregulates the expression of ALK5 and competitively inhibits its kinase activity, thereby suppressing the TGF-β signaling pathway via SMAD and non-canonical cascades. This mechanism underpins its ability to ameliorate BLM-induced pulmonary fibrosis in mice and inhibit TGF-β1-induced epithelial cell migration, EMT, and fibroblast transdifferentiation. Notably, DHM’s effective concentrations in human cells (A549 and MRC-5) were ten-fold lower than those reported in primary mouse fibroblasts [
77], suggesting superior efficacy in human systems.
In the experimental pulmonary fibrosis mouse model induced by BLM, the inflammatory response predominates within the first 7 d post-intratracheal BLM administration, accompanied by epithelial cell proliferation, resembling acute lung injury. Between days 7 and 10, inflammation subsides, ECM deposition begins, and fibrotic lesions emerge, persisting for 3-4 weeks [
78]. Consequently, our study initiated treatment 7 d post-BLM induction. To monitor fibrosis progression, mice were sacrificed at multiple time points for the purpose of histopathological analysis. No mortality was observed during the initial 7 d, excluding acute lung injury as a cause of death. From day 14 to day 28, BLM-treated mice showed marked thickening of alveolar walls, disorganized tissue architecture, along with inflammatory cell invasion and collagen build-up in lung tissue (Figs. S7(a) and (b) in Appendix A). Ashcroft’s fibrosis scoring confirmed severe and progressively worsening fibrosis in BLM-treated mice (Fig. S7(c) in Appendix A). While Li et al. [
77] showed DHM’s therapeutic effects in BLM-induced experimental pulmonary fibrosis, their study employed one high dose (300 mg·kg
-1) and lacked a reference drug for efficacy comparison. In contrast, our study evaluated DHM at lower-level doses (25 and 50 mg·kg
-1). Additionally, our study included PFD (50 mg·kg
-1) as a positive control. DHM treatment significantly reduced body weight loss and mortality compared to PFD. Histologically, both DHM and PFD mitigated inflammation and delayed fibrosis progression. Quantitative analyses further revealed that DHM achieved efficacy comparable to PFD in improving respiratory function and reducing inflammation during fibrosis. Collagen deposition, a hallmark of excessive ECM production, was substantially repressed in DHM-treated mice, reflected by decreased hydroxyproline content in lung tissue. Collectively, these data indicate that DHM exerts significant therapeutic effects at lower doses and is not inferior to the marketed antifibrotic drug PFD under the experimental conditions of this study.
Notably, with comparable efficacy to PFD, the distinct advantage of DHM lies in its wider safety window. At equivalent doses, DHM exhibits significantly lower toxicity than PFD and remains well-tolerated even at 2-20 times the therapeutic dose. Mice treated with DHM showed no abnormal changes in body weight or signs of liver and kidney function impairment (Fig. S5), confirming its excellent safety profile. Furthermore, DHM’s superior hydrophilicity facilitates the development of oral formulations for large-scale production, while also enhancing gastrointestinal absorption and target tissue penetration to optimize bioavailability. Mechanistically, DHM acts with greater precision by directly targeting ALK5, the core receptor of the TGF-β pathway, thereby minimizing off-target effects. This contrasts with PFD, which features a broader, multi-pathway mechanism of action [
79-
84].
In addition to the benefits of DHM compared to PFD, a review of existing literature suggests that DHM may also exhibit certain unique advantages over other anti-fibrotic agents in clinical stages, such as BI 1015550 and TRK 250. Compared to the oral phosphodiesterase 4B (PDE4B) inhibitor BI 1015550, which has completed Phase III trials for IPF [
85], DHM offers a more direct therapeutic strategy. Rather than indirectly modulating inflammation and fibrosis by regulating cyclic adenosine monophosphate (cAMP) degradation, DHM directly inhibits the core driver of fibrotic signaling, potentially yielding superior specificity and efficiency [
86]. Additionally, DHM is less frequently associated with the gastrointestinal side effects (e.g., nausea and diarrhea) commonly linked to PDE4 inhibitors, suggesting a more favorable tolerability profile [
87,
88]. In contrast to TRK 250, an inhaled nucleic acid drug that has completed Phase I trials [
89], DHM possesses several advantages inherent to its identity as an oral small molecule. These include superior stability, more convenient production and storage, likely higher long-term patient adherence, a well-defined ALK5-targeted mechanism, and robust preclinical safety data. Its more affordable cost also deserves emphasis. Taken together, these attributes highlight DHM as a considerably promising candidate for the treatment of pulmonary fibrosis.
In this study, DHM was initially identified as a potential inhibitor of TGF-β/SMAD signaling pathway via an ML-based classification model. This screening strategy prioritized the canonical TGF-β/SMAD signaling pathway due to its validated role in IPF pathogenesis [
5,
8]. Subsequent experiments demonstrated DHM’s ability to inhibit the phosphorylation of SMAD-dependent and non-canonical TGF-β signaling pathway-related mediators in cell and animal models (
Figs. 4,
5(i), and
5(k)). Based on these findings, we reasonably hypothesized DHM would ameliorate IPF by down-regulating the crucial mediator of SMAD-dependent and non-canonical TGF-β signaling pathway. ALK5, a kinase receptor for TGF-β1, holds a central position in TGF-β signal transduction. It mediates the phosphorylation and activation of itself and downstream effector proteins [
23]. This pivotal role renders ALK5 a promising therapeutic target for fibrosis. Previous studies have shown that inhibiting ALK5 kinase activity delays the progression of pulmonary and liver fibrosis [
90,
91]. Notably, AGMB-447, a small-molecule ALK5 inhibitor, has recently been approved as orphan drug for IPF by FDA and currently in Phase I clinical trial [
92]. Leveraging these established insights, our hypothesis is that DHM exerts anti-fibrotic effects through direct interaction with ALK5. By inhibiting ALK5 kinase activity, DHM curbs the TGF-β1-activated pulmonary fibrosis signaling axis, providing a potential treatment for experimental pulmonary fibrosis. This hypothesis aligns with the growing interest in targeting ALK5 for fibrosis therapy and underscores the therapeutic potential of DHM.
To elucidate the molecular mechanism of DHM’s action on ALK5, we employed an integrated approach combining biophysical, computational, and biochemical techniques. The initial validation of DHM-ALK5 interaction was achieved through WB-CETSA, harnessing ligand-induced protein thermal stabilization principle [
93,
94]. Molecular docking simulations predicted DHM’s binding to the intracellular kinase domain of ALK5 with a binding energy of -9.60 kcal·mol
-1, while molecular dynamics simulations and MM/PBSA calculations confirmed stable binding (free energy = (-35.20 ± 0.17) kcal·mol
-1). Experimental validation using MST revealed a
Kd of (5.09 ± 0.20) μmol·L
-1, demonstrating comparable affinity to the classic ALK5 inhibitor SB431542 (
Kd = (11.25 ± 2.93) μmol·L
-1; Fig. S8 in Appendix A). Further experiments revealed that DHM competitively inhibits ALK5 kinase activity by binding to its catalytic domain. This is the first time to identify that DHM acted as a novel ALK5 kinase inhibitor with therapeutic potential for pulmonary fibrosis (
Fig. 7). Unlike most tyrosine kinase inhibitors, which often exhibit cardiotoxicity, DHM showed no myocardial toxicity at tested concentrations (Table S3 in Appendix A). Western blotting assays further demonstrated that DHM also reduces ALK5 protein expression on the cell membrane, which indicated that DHM represses the TGF-β/ALK5 signaling cascade by the dual mode of action, reducing the ALK5 expression and inhibiting the kinase activities.
This study highlights the transformative role of AI and ML in drug research and discovery. By efficiently screening thousands of natural compounds, we identified DHM as a novel ALK5 inhibitor with significant antifibrotic effects. This approach not only accelerates the discovery of TGF-β pathway inhibitors but also opens avenues for exploring other signaling targets. Among the untested predicted candidates, additional potent inhibitors may await discovery, further advancing IPF therapeutics.
5. Conclusions
To conclude, through a combination of an ML-enabled classification framework and a dual-luciferase reporter assay, our research characterized DHM as a novel inhibitor of the TGF-β/ALK5 signaling axis. Mechanistically, we reported for the first time that DHM reduces the expression of ALK5 on cell membranes and competitively impedes its kinase function. Consequentially, the TGF-β/ALK5 signaling cascade is dampened, leading to the attenuation of experimental pulmonary fibrosis in mice. Additionally, DHM effectively inhibits TGF-β1-triggered epithelial cell migration, EMT, and fibroblast transdifferentiation, highlighting its potential as a therapeutic agent against IPF. Compared to the existing ALK5 inhibitors [
12], DHM exhibits a higher degree of structural innovation, substantially broadening the chemical space of the ALK5 inhibitors. Furthermore, DHM is derived from natural products and contains advantages such as low toxicity and good water solubility, making it a more promising candidate for the research and development of novel anti-IPF drugs, with the potential to enhance the medical treatment of IPF.
PDF
(5833KB)
