AlphaFold transformed structural biology, but its impact on drug discovery has been slower and narrower. It gave researchers free predicted structures for more than 200 million proteins and won a Nobel Prize. Yet no approved drug owes its existence to AlphaFold, because protein structure was rarely what made drugs fail.
In November 2020, at the CASP14 protein structure prediction competition, DeepMind’s AlphaFold 2 predicted the shapes of proteins with accuracy close to laboratory methods. Problems that had taken structural biologists months or years of crystallography could now be approximated in hours on a computer.
Five years later, the numbers are extraordinary by any standard in science. Google DeepMind reports that the AlphaFold database, built with EMBL’s European Bioinformatics Institute, holds more than 200 million predicted structures and has been used by over 3 million researchers in more than 190 countries, including over a million in low- and middle-income countries. In 2024 Demis Hassabis and John Jumper shared the Nobel Prize in Chemistry for the work, alongside David Baker for computational protein design.
And yet, if you ask how many medicines on pharmacy shelves exist because of AlphaFold, the honest answer is none so far.
Both facts are true, and the gap between them is the most useful thing to understand about AI in drug discovery. AlphaFold solved a problem that was a bottleneck for biology. It was only occasionally a bottleneck for drug pipelines. Drugs mostly fail in people because the target was wrong or the molecule was toxic, and knowing a protein’s shape, while valuable, answers neither question directly.
What did AlphaFold actually change?
The biggest change was access. Before AlphaFold, a protein structure was a scarce and expensive thing. You needed a crystallography or cryo-electron microscopy lab, a protein that could be purified and coaxed into shape, and time. Well-funded labs working on well-studied proteins had structures. Everyone else guessed.
After July 2021, when DeepMind published AlphaFold 2 in Nature and released the code, and especially after the database expanded to cover almost every catalogued protein in 2022, a structural hypothesis became something any researcher could download. That shifted work in basic biology dramatically. DeepMind’s five-year review cites independent analysis suggesting that researchers using AlphaFold submit more novel experimental structures, and that AlphaFold-linked research is more likely to be cited in clinical articles and patents than typical structural biology work.
For neglected diseases and under-resourced labs, this matters more than for large pharma. A parasite protein that no company would ever pay to crystallize now has a predicted structure. Whether anyone pays to develop a drug against it is a separate question, one that the market answers the same way it always has, as our piece on the biggest killers getting the smallest share of aid describes.
Why hasn’t AlphaFold produced a drug yet?
Three reasons, in rising order of importance.
The first is time. Drug development takes a decade or more. A target first explored with an AlphaFold model in 2021 would, on a normal timeline, be in early clinical trials now at best. Judging the tool on approvals in 2025 is premature.
The second is that a predicted structure is not the same as a druggable one. AlphaFold 2 predicts a single, static shape, usually without the small molecule a drug designer cares about. Proteins move. Binding pockets open and close. Drugs often bind to shapes the protein adopts only briefly. The evidence on using AlphaFold 2 models for virtual screening, where millions of compounds are computationally docked against a target, is mixed. A retrospective benchmark across 22 targets found AlphaFold models consistently performed worse than experimental structures across four docking programs, and other studies found predicted structures behave more like empty, unbound protein structures than like the ligand-bound shapes that screening works best with.
The third, and the one that matters most, is that structure was rarely the reason drugs failed. For many established drug targets, experimental structures already existed. The expensive failures in pharma happen in phase 2 and phase 3, when a drug that binds its target beautifully turns out not to change the disease, or causes harm nobody predicted. A better structure helps design a better key. It does not tell you whether you picked the right lock.
Is AlphaFold 3 different for drug discovery?
Somewhat. AlphaFold 3, described in Nature in May 2024, predicts how proteins interact with other molecules, including small-molecule ligands, DNA and RNA. That moves the tool closer to the question drug designers actually ask: not “what does this protein look like?” but “how will this compound sit in it?”
It also changed the commercial picture. AlphaFold 3’s code was released for academic use in late 2024, but the most advanced drug design work built on it sits with Isomorphic Labs, the Alphabet company spun out of DeepMind. Isomorphic signed discovery partnerships with Eli Lilly and Novartis in early 2024, reported to carry up to roughly $3 billion in combined milestones, and raised $600 million in March 2025 in a round led by Thrive Capital. The company has said it is preparing to take its first programs into clinical trials.
That is the right place to watch. If AlphaFold’s descendants are going to change drug pipelines rather than biology papers, the evidence will show up in Isomorphic’s and its partners’ clinical results. Until those exist, the commercial case rests on partnership terms, which are mostly milestones that may never be paid.
Where has AlphaFold already helped find molecules?
There are real, if early, examples. Insilico Medicine used an AlphaFold-predicted structure of CDK20, a liver cancer target with no experimental structure at the time, to find a first small-molecule inhibitor, reported in Chemical Science in 2023. A prospective 2024 screening study against TAAR1, a receptor involved in psychiatric conditions, reported a hit rate of about 60 percent from an AlphaFold model, more than twice the rate from a conventional homology model.
Those results are interesting partly because they cut against the retrospective benchmarks. When researchers actually test compounds in a lab, AlphaFold models sometimes do better than their benchmark scores predict. The field has not yet settled why, and that uncertainty is itself informative. The tool is powerful enough to use, and not yet understood well enough to trust blindly.
A worked example: where the months actually go
Consider a hypothetical early discovery program at a mid-sized biotech, with round numbers for illustration only.
Before AlphaFold, the team spends around a year trying to obtain an experimental structure of a novel target, fails to crystallize it, and builds a rough homology model instead. Screening against that model produces a handful of weak hits. Medicinal chemistry takes another two years to reach a candidate.
With AlphaFold, the team has a plausible structure on day one. Screening starts immediately. Perhaps the first hits come a year earlier, and chemistry is somewhat faster because the team has a better picture of the pocket.
Now follow the candidate forward. Preclinical toxicology, manufacturing, phase 1, phase 2 and phase 3 take, together, most of a decade and the overwhelming majority of the money. AlphaFold saved a year in a process that takes twelve or more and changed nothing about the probability that phase 2 succeeds. That is a real gain. It is not a transformation of the economics of medicine.
This is also why the most valuable use of AlphaFold may be killing bad ideas early. A structure showing that a target has no plausible binding pocket saves a program before it spends anything serious. Those saved costs never appear in a press release.
The strongest case that AlphaFold’s impact is underrated
The counterargument is serious. It holds that judging AlphaFold by drug approvals misunderstands what it is. AlphaFold is infrastructure, like the Human Genome Project or the Protein Data Bank that trained it. Nobody asks how many drugs the genome “discovered.” It changed the ground every discovery stands on.
On that view, AlphaFold’s drug discovery impact is already everywhere and simply unattributed. Thousands of target validation decisions, mechanism hypotheses and experimental designs have used predicted structures. Some of those will become drugs, and nobody will label them “AlphaFold drugs.”
This is largely right, and it is a better frame than the hype. But it concedes the key point. If AlphaFold is infrastructure, its value is diffuse and slow, and it should not be sold, as it often has been in investor decks, as a direct route to faster approvals. Infrastructure makes everyone somewhat better. It rarely makes one company decisively better, which matters for how much investors pay for “AlphaFold-powered” platforms, a theme we touched on in health venture capital in 2025.
What this means for biotech founders, pharma leaders and funders
For founders, AlphaFold access is no longer a differentiator. It is free and universal. A startup whose pitch is “we use AlphaFold” is pitching a tool every competitor has. Differentiation lives in proprietary experimental data, dynamics, binding measurements and, above all, target selection.
For pharma leaders, the practical value is in early-stage triage and in targets that were previously structurally dark. The discipline is to treat predicted structures as hypotheses to be tested, not answers. The evidence that predicted structures sometimes mislead docking is strong enough that any program relying on one should confirm key binding assumptions experimentally before large chemistry investments.
For research funders and governments, AlphaFold is a case study in the returns to open scientific infrastructure. It was trained on decades of publicly funded structures in the Protein Data Bank. Its largest measured effect has been on academic biology worldwide, including in lower-income countries. The policy lesson is to keep funding the unglamorous public datasets that make the next AlphaFold possible, and to fund experimental validation alongside prediction.
For clinicians and patients, AlphaFold is not yet something that changes treatment decisions. Its payoff, if it comes, will arrive through drugs that go through ordinary trials, judged by ordinary evidence.
Frequently asked questions
Has AlphaFold led to any approved drugs?
Not as of late 2025. AlphaFold has been used in many discovery programs, and companies such as Insilico Medicine have reported inhibitors found with AlphaFold-predicted structures. But drug development takes a decade or more, and no approved drug is publicly attributed to AlphaFold. Isomorphic Labs, the Alphabet spinout built around its successors, has said it is preparing for its first clinical trials.
What is the difference between AlphaFold 2 and AlphaFold 3?
AlphaFold 2, published in 2021, predicts the 3D structure of individual proteins with high accuracy. AlphaFold 3, published in Nature in May 2024, predicts how proteins interact with other molecules, including DNA, RNA and small-molecule ligands. That makes AlphaFold 3 more directly relevant to drug design, which depends on how a compound binds to its target.
Can AlphaFold structures replace experimental structures in drug design?
Not reliably. Some prospective studies found AlphaFold models produced good screening hit rates, but benchmark studies have found them weaker than experimental structures for docking. Predicted structures are static and often lack the shapes proteins adopt when drugs bind. Most teams treat them as strong hypotheses and confirm critical details with experiments.
Who can use AlphaFold?
The AlphaFold Protein Structure Database, run with EMBL-EBI, is freely available to anyone, and Google DeepMind reports more than 3 million users in over 190 countries. AlphaFold 3’s code and model weights were released for academic, non-commercial use. Commercial drug design using the most advanced versions is largely done through Isomorphic Labs and its pharma partners.
The real impact
AlphaFold’s legacy in medicine will probably be enormous and hard to see. It will show up as better-chosen targets, abandoned dead ends and faster early chemistry, spread across thousands of programs, most of which will never mention it. That is what good infrastructure looks like.
What it will not do is make phase 2 trials succeed. That still depends on biology in people, and no structure prediction, however beautiful, can run that experiment for us.
Related on the GHC blog: AI Diagnostics: The Evidence Is Strongest Where the Question Is Narrowest and Fast Approval Isn’t the Danger. Slow Withdrawal Is..
Has AlphaFold changed a decision in your own research or pipeline, or is it still mostly a tool for papers? We’d like to hear where it has earned its place.
