AI could accelerate drug discovery. But only if we can trust it.

One of the most consequential advances in artificial intelligence isn’t an eerily conversational chatbot—it’s a new way to unpack the unique 3D structures of proteins. This powerful deep-learning algorithm, dubbed AlphaFold, turns a process that once took scientists years to complete in the lab into a computer program that could run in less than an hour.

The implications for medicine are immense: once the molecular nuances of a protein’s structure have been identified, researchers can begin to target it with drugs, correcting dysfunctions, combating infections, and improving health. But before AI can transform biomedicine, researchers will need to demonstrate that the algorithm’s predictions are accurate enough for drug developers to confidently use them to guide discovery.

A new paper in Science suggests this may be the case. When researchers used sophisticated software to sift through billions of compounds—searching for potential new drugs by matching them against protein structures—they found that structures predicted by AlphaFold did indeed have the same ability to guide virtual drug screening as experimental structures, at least for the two drug targets they examined.

The findings are among the first to demonstrate that one iteration of this AI technology, AlphaFold2, can be effectively used in this manner. “Until now, studies have suggested that AlphaFold2 is worse than experimental structures for structure-based drug screen tasks,” says Jiankun Lyu, first author on the paper, who conducted much of the research at University of California, San Francisco before joining Rockefeller to complete the project. “We found, in the drug targets we tested, that the algorithm’s model is as reliable as experimental structures, when used as inputs in our program to discover ligands, which are the binding molecules you need to identify for drug discovery.”

We sat down with Lyu to discuss the promise of the latest version of the technology, AlphaFold3, the limitations of deep learning, and what it all means for drug discovery.

What does your paper tell us about AlphaFold’s potential for advancing medicine?

Our expectation, based on prior work, was that AlphaFold would be worse than experimental methods at structure-based ligand discovery.

We were surprised to find that, when analyzed prospectively, AlphaFold’s predicted structures are sometimes close enough to structures obtained experimentally. We estimate that, in approximately one-third of cases, an AlphaFold-predicted structure could significantly expedite a project. The potential to accelerate project timelines by up to a few years, compared to obtaining a new structure through experimental methods, represents a substantial advantage.

How will AlphaFold3 improve upon this?

On one hand, AlphaFold3 is a huge upgrade from AlphaFold2. The prior model could only predict single-chain protein structures; only with their Multimer add-on could AlphaFold2 predict some protein complexes. But the newest model can predict post-translational modification and the small molecule protein complexes. Put simply, the developers claim that the AI can now forecast protein-molecule complexes involving DNA, RNA, and other molecules.

The problem is that the most recent release is a black box.

When AlphaFold2 was first released, the team released their model as well. There was no real limitation on how many proteins a user could predict. As a consequence, we were able to interrogate the algorithm and broader applications in basic science and drug discovery, as in our most recent paper. The latest model, unfortunately, is only available on a server—they are not releasing the model—and the number of structures that can be predicted per day is limited. There are some signs that they may change this policy and increase transparency in the next six months. But if they don’t open the model up to academic screening use, our present study will be the last of its kind. We would not have been able to run the current study on AlphaFold3. And without that, we can’t know whether the new model is better for templating drug discovery.

Does this shift in policy make you less optimistic about the future of AI and medicine?

I, personally, am enthusiastic! But I’m advising caution because a lot of AI is currently overpromised and under-delivered. I’m concerned that, if we don’t treat it carefully now, AI in biomedicine will fizzle out and end up being just another hype. That could set us back decades.

So the future remains bright?

Absolutely. This is one of the hottest research areas, and there’s a huge market for accurately predicting protein complexes in both basic research and industry. In the lab, we need 3D models of the complexes we’re interested in investigating to elucidate crosstalk in many mechanistic studies, and on the industry side, the more accurate and easy to obtain these models are, the more researchers can start imagining antibody and nanobody biologics or small molecule drugs that interact with therapeutic targets. Although that’s not all it takes to make a drug, getting an accurate model is a crucial early step that also guides further drug optimization.

There were once many people who didn’t think AI and deep learning models would be able to do these sorts of things. We still aren’t sure it can—but it is looking more and more likely.

Besides increased transparency from AI companies, what will it take to improve deep learning models so that it becomes a practical tool for drug discovery?

Many studies have shown that AI is capable of doing great things for biomedicine, but how well it can do those things is bottlenecked by the availability of experimental data to train the AI.

Where AI is already succeeding happens to be in those areas in which basic science has generated a lot of data experimentally. So now that we have many AI architectures, we need to go back to the bench and generate more high-quality data, to feed these data-hungry algorithms until they produce better predictions. That’s when the breakthroughs will come.