AlphaFold Three Years Later: What the Structure Revolution Actually Changed

AlphaFold 2’s release in 2021 produced one of those rare moments in science where the field’s reaction was not skepticism but something closer to vertigo. Researchers who had spent careers solving one protein structure per PhD had just watched a model accurately predict structures for nearly every protein known to science. The AlphaFold Protein Structure Database eventually grew to cover over 200 million proteins. Every organism with a sequenced genome, every protein that protein codes for, solved.

The question was never really whether the structures were accurate. They were — with some important caveats about disordered regions and multi-chain complexes that took another two years to substantially resolve. The question was what having all these structures would actually enable. That question has now had three years to start answering itself, and the answer is more complicated than either the triumphalists or the skeptics predicted.

The Drug Discovery Gap

The clearest and most expensive expectation was pharmaceutical. If you know the shape of every disease-relevant protein, the reasoning went, you can rationally design drugs to interact with it. Structure-based drug discovery had been a successful methodology for decades — HIV protease inhibitors, kinase inhibitors for cancer, the whole class of structure-guided fragment-based drug design. AlphaFold, the logic went, would democratize this by eliminating the bottleneck of experimental structure determination.

By early 2027, the scorecard for AI-assisted drug discovery is real but decidedly mixed. Structure-based AI tools have contributed meaningfully to lead identification — finding candidate molecules that fit a protein’s binding pocket — and have compressed some early-stage timelines. Insilico Medicine’s INS018_055, a drug for idiopathic pulmonary fibrosis that began as an entirely AI-designed candidate, entered Phase II trials in 2024 and remains in development. Recursion Pharmaceuticals has a pipeline that uses AI-derived structural information as one input among many. Isomorphic Labs, DeepMind’s drug discovery spinout, has disclosed early-stage collaborations with Eli Lilly and Novartis, though the therapeutic candidates are years from any public validation.

The hard constraint that structures cannot eliminate is the gap between binding and efficacy. A molecule that fits neatly into a protein’s binding pocket may still fail for dozens of reasons: it might not cross cell membranes, might be metabolized too quickly, might bind the right target in a test tube and the wrong one in a body. Of the roughly 1,400 small molecule drugs currently in clinical trials globally, not one has yet been approved on the basis of a structure predicted by AlphaFold — the modeling contributed to early stages, but every candidate has a long road of experimental chemistry and biology behind it. That road has not shortened.

Where the Revolution Actually Landed

The honest beneficiaries of AlphaFold, as of early 2027, are not the drug companies. They are the researchers studying obscure proteins nobody had ever crystallized because they were too difficult, too unstable, or simply not commercially interesting enough to justify the effort.

Structural biology, before AlphaFold, had a massive sampling bias. The Protein Data Bank contains approximately 220,000 experimentally solved structures. Almost all of them are from proteins that are abundant, stable, and scientifically or commercially interesting — meaning evolutionarily conserved proteins from model organisms and drug targets from humans. Entire branches of the tree of life had no structural coverage. The proteins of archaea, of extremophile bacteria, of organisms living in deep ocean hydrothermal vents — basically unknown at the structural level.

That is now changed. A 2025 paper in Science used AlphaFold structures to analyze the structural diversity of microbial proteins from metagenomic datasets — sequences pulled from environmental samples with no cultured organism attached. The analysis revealed entire protein families with no previously known structural homologs. Not new folds — there are only perhaps 1,300 distinct protein folds in existence — but new arrangements, new functional contexts, new evolutionary lineages.

The enzyme engineering community has been perhaps the most practically productive user of the new structural resources. Directed evolution, the technique that won Frances Arnold the Nobel Prize, works by randomly mutating proteins and selecting for improved function. Structure-guided directed evolution uses structural information to make the mutations less random — to target positions that matter for the desired property. AlphaFold structures, especially for non-model organisms with interesting biochemistry, have materially accelerated this work. A set of novel esterases for degrading plastic polymers, an oxidase for a pharmaceutical intermediate synthesis that would otherwise require protecting group chemistry, several thermostable variants of commercially useful enzymes — these are small stories, not headline stories, but they accumulate into something real.

The Dynamics Problem

The main scientific limitation of AlphaFold is one that was apparent from the beginning but that has resisted solution for longer than many expected: it predicts static structures.

Proteins are not static. They breathe. They shift conformation in response to binding partners, to post-translational modifications, to pH, to membrane context. Many proteins have two or more distinct conformational states that are biologically meaningful — and the difference between them, sometimes just a few angstroms of movement in a key loop, is often the difference between active and inactive, open and closed, signaling and silent.

AlphaFold 3, released in 2024, made substantial progress on protein-ligand and protein-protein complexes. It did not solve dynamics. The problem requires not just structure prediction but free energy sampling — understanding the landscape of conformations that a protein explores over time — which remains computationally expensive enough that running it at proteome scale is not practical. OpenFold, ESMFold, and several academic groups have made progress on faster approximate dynamics prediction, but the fundamental problem of capturing conformational ensembles remains open.

This matters clinically because many important drug targets are “disordered” proteins — proteins, like tau in Alzheimer’s or alpha-synuclein in Parkinson’s, that don’t have a single stable structure at all. Intrinsically disordered proteins have resisted structural biology for decades specifically because they have no stable structure to solve. AlphaFold handles them poorly. The next frontier of structural biology is not more structures; it is better models of motion.

The Quiet Shift in How Research Gets Done

The less dramatic but perhaps more durable impact of AlphaFold is methodological. It has changed what questions biologists think are worth asking.

Before 2021, a graduate student interested in the structure of an unstudied bacterial protein faced a multi-year project of cloning, expression, purification, crystallization, and X-ray data collection just to get a structure — and there was no guarantee the protein would crystallize at all. That bottleneck shaped which research got proposed. Structural work required structural expertise, expensive equipment, and a supportive core facility. It was not casual.

Now a student can get an AlphaFold prediction in minutes. The prediction may need experimental validation if the structural details matter precisely — and for drug discovery, they often do. But for generating hypotheses, for comparing a novel protein to its evolutionary relatives, for asking whether a mutation you’re interested in affects the active site, the prediction is usually sufficient. The activation energy for thinking structurally has dropped to nearly zero.

This is a real shift. It is hard to measure in publications because it shows up as a change in research texture — structural information appearing as context in papers that would not previously have had any structural content, comparative genomics studies that incorporate structural data, evolutionary analyses that annotate function based on predicted structures. The proteome-scale thinking that was previously possible only for groups with specific computational infrastructure is now available to anyone with a university account.

AlphaFold solved a problem that the field had defined for itself. It turned out that the problem the field had defined, while important and real, was not the whole problem. This is not a failure; it is how science works. You solve the problem you can state, and the solution reveals the shape of the problems you couldn’t state before. The next decade of structural biology is going to be shaped by questions that AlphaFold made it possible to ask.