A drug developer trying to devise a new cancer treatment, a virologist investigating how a virus invades cells, an evolutionary biologist probing the effects of mutations on an organism’s fitness—these are just some of the scientists who need to know the three-dimensional structures of specific proteins. However, deducing a protein’s shape through experiments is laborious and costly [
1]. By the early 2020s, researchers had succeeded with only about 200 000 of the molecules, less than 0.1% of the total that known organisms produce [
2]. But in mid-2022, scientists announced that an artificial intelligence (AI) model known as AlphaFold, developed by the London, UK-based Google subsidiary DeepMind, had predicted the structures of almost all of the more than 220 million unsolved proteins (
Fig. 1) [
3], [
4]. Experimentally determining the structures of 200 000 proteins took researchers about 50 years and cost around 2 billion USD, said Marc Zimmer, professor of chemistry at Connecticut College in New London, CT, USA. “AlphaFold can do that in a couple of days for essentially nothing because it is online.”
AI is not only illuminating the architecture of existing molecules. It is also inventing new ones. Several AI-designed drugs have reached clinical trials, and for the first time one has progressed to a phase II trial that will test whether it works [
5]. And AI is delivering news-worthy discoveries in other areas of research, including geology [
6], physics [
7], neuroscience [
8], and math [
9]. AI “is making connections we have never made before. That is the breakthrough,” said Zimmer.
But some researchers worry about the increasing reliance on AI in science. AI can mislead. It concocts false results, a problem known as hallucination [
10], [
11]. And even accurate findings may not be scientifically relevant. Like all scientific discoveries, those made by AI need to be confirmed. But many researchers are skimping on this step, said Casey Bennett, assistant professor of computing and digital media at DePaul University in Chicago, IL, USA. “A lot of people are not following the scientific process to do it correctly to verify the answer.” Moreover, researchers may not be able to validate some of AI’s results because its procedure for drawing conclusions is often obscure, a problem that some experts worry could spark a “reproducibility crisis” in science [
12], [
13].
AI encompasses several technologies, including robotics. But researchers often use the AI variety known as machine learning, in which computers draw inferences or identify patterns by poring over vast data sets [
14]. “Machine learning is good at picking out patterns that are beyond our cognitive capabilities,” said Bennett. That power has already made AI indispensable for several fields, including astronomy. Without help from machine learning, astronomers could not hope to analyze the huge amounts of data generated by telescopes and other instruments, said Chris Impey, professor of astronomy at the University of Arizona in Tucson, AZ, USA. For example, he noted, the Vera C. Rubin Observatory in Chile, which is set to begin a ten-year survey of the southern sky in 2024, will produce up to 40 TB of data every night. AI’s ability to detect anomalies makes it particularly valuable, he said, allowing researchers to identify novel objects that might otherwise go unnoticed. “In astronomy, you do not always know what you are looking for.” Astronomers have already used AI to pinpoint a new planet outside our Solar System, reveal an asteroid that could potentially strike Earth, and make other discoveries [
15], [
16].
As noted above, researchers trying to crack one of the toughest problems in biology—predicting how proteins fold into complex shapes—have also received a game-changing boost from AI [
4]. A protein’s structure dictates how it works but predicting a protein’s shape from only its sequence of amino acids, or chemical building blocks, is extremely difficult [
17]. Although researchers have been testing the predictive abilities of AI models—including a previous version of AlphaFold—for decades, it was not until a 2020 protein folding contest, the Critical Assessment of Protein Structure Prediction, that the technology showed its prowess. AlphaFold beat the more than 50 other competitors by a large margin, and two-thirds of its predictions were as accurate as experimentally determined structures—a significant jump in performance for AI models [
18], [
19].
DeepMind scientists had trained the model by feeding it the known structures of about 170 000 proteins. When AlphaFold is tasked with predicting the shape of an unsolved protein, it uses these structures as a guide to determine which amino acids might be near each other [
18]. The model’s developers improved on the previous version of AlphaFold by adding an algorithm that builds a protein’s structure by starting with small pieces and moving on to larger and larger ones [
18], [
20].
AlphaFold’s success set off a rush to apply it. More than 400 000 people accessed the AlphaFold database in the first nine months after it went online, and the number of papers that referenced the model more than quadrupled in the same time [
21]. Researchers have enlisted AlphaFold to uncover a potential new drug for treating liver cancer [
22], predict the structure of a key protein from a dangerous virus [
23], and generate the most comprehensive model of the nuclear pore complex, the elaborate molecular channel that controls what enters and exits the nucleus of a cell [
21]. “We can suddenly do things really easily that were extremely difficult or impossible before,” said Zimmer. DeepMind has also released an upgraded version of AlphaFold that can forecast how proteins will interact with other types of molecules, a capability that could benefit drug developers [
24].
Some news articles declared that AlphaFold had “solved” the protein folding problem, but AI models do not deliver the final word on molecular structure, cautioned Tamir Gonen, professor of biological chemistry at the University of California, LA, USA. AlphaFold often gets the overall structure of proteins right, he said. That capability has made it easier for scientists studying specific proteins to obtain models of what the molecules might look like, and that “may help them explain their experiments,” Gonen said. “I choose my words carefully,” he added, because predicted structures from AlphaFold can also be wrong. “You have to validate these models. If you do not and take them as truth, you could have a big problem,” said Gonen. In short, he said, AI is not ready to take over from biologists. “There may come a time when we no longer need to solve structures experimentally, but we are not there yet.”
AlphaFold also struggles with certain protein features, Julie Forman-Kay, senior scientist at the Hospital for Sick Children Research Institute in Toronto, Canada, and other researchers have found. All human proteins contain sections, known as intrinsically disordered regions, that do not fold into a definite shape and instead continually shift among different conformations [
25]. More than one-third of the amino acids in proteins are in disordered regions, and proteins range from having almost no disorder to being completely disordered [
26]. In 2023, Forman-Kay and her team analyzed how AlphaFold handles intrinsically disordered regions. Researchers had thought that AlphaFold would mark its predictions for these regions as uncertain, thus warning users that the putative structures were possibly incorrect. But about 15% of the time, the model assumes that these regions will fold into specific shapes and gives its predictions a high level of confidence, Forman-Kay and her colleagues reported [
27]. In most of these cases, the disordered region likely adopts a particular shape under certain conditions, such as when it binds to another protein, said Forman-Kay. But those conditions are unknown, and many times disordered regions fold into different structures under different conditions, making AlphaFold’s predicted structures inaccurate, she said. Few scientists are aware of these limitations, Forman-Kay said, and they may mistakenly conclude that the model’s predictions are definitive. She still believes that AI provides important insights into protein folding, but researchers need to recognize that “it is just a tool” and needs to be applied cautiously.
AI does not just help researchers analyze data. In fields such as space science and astrophysics, it can also help decide what data scientists see. Today, spacecraft and observational satellites can gather much more data than they have the bandwidth to transmit back to Earth. That is why more and more space missions also include on-board computers equipped with AI modules. These AI systems filter out bad or low-quality data, allowing the craft to send only the high-quality data back to Earth [
28]. Giving AI algorithms that power also places a large burden on them, said Simon Wing, a physicist at Johns Hopkins University Applied Physics Laboratory in Laurel, MD, USA. “We need to be careful if we use [AI] to reduce data,” he said. If the AI is not trained or applied properly, “it may misclassify useful data as bad. If that happened, the data would be lost forever.”
Bennett worries that many of the published results produced by AI models are wrong or irrelevant. Researchers are not taking care to double-check that their findings are accurate and scientifically informative, he said. “The problem is not, ‘Can I find a pattern?’ That is easy. The problem is finding out which patterns are meaningful in the real world. That is hard.” The underlying cause, according to Bennett, is that researchers who do not understand AI’s subtleties and limitations are rushing to use it. He thinks that before scientists apply AI in their research, they should have to pass a certification course set up by a professional body like the organizations that license lawyers and doctors. “Machine learning, just like surgery, requires the proper level of training and experience,” he said. “Otherwise, we can produce misleading answers and bad solutions.”
Other researchers do not go that far. But some scientists who acknowledge AI’s constraints are working to give its results more credence. AI’s co-called black box problem—how the models arrive at their conclusions is often mysterious—has long been a concern [
29]. Even some of AlphaFold’s “thinking” is inexplicable, the chief executive officer of DeepMind has admitted [
30]. To overcome the problem, researchers are developing so-called explainable AI, models that follow specific rules and procedures so that their reasoning is transparent [
31].
That is just one step that could make AI results more credible, but scientists also need a clearer understanding of what the models can and cannot achieve, experts say. AI tools can be illuminating, said Wing, “as long as they are used correctly, and their limitations are acknowledged.”
PDF
(384KB)
