AlphaFold Three Years Later

When DeepMind released AlphaFold 2’s predictions for nearly the entire human proteome in July 2021, the response from structural biologists ranged from astonishment to mild existential crisis. John Moult, who had organized the Critical Assessment of Protein Structure Prediction competition since 1994 — the biennial contest that measured progress on the protein folding problem — said it was, in his phrasing, like a line being crossed that the field had expected to approach gradually over decades. What arrived instead was a step function.

The protein folding problem had been an organizing challenge in molecular biology since Christian Anfinsen won the Nobel Prize in 1972 for demonstrating that a protein’s three-dimensional structure is determined by its amino acid sequence. The challenge: given the sequence, predict the structure. Given the structure, understand the function. Given the function, design drugs that interact with it. For 50 years, the prediction step was the bottleneck. Experimental structure determination required years of work, expensive equipment, and often failed entirely for proteins that resisted crystallization or couldn’t be expressed at sufficient purity. AlphaFold removed the bottleneck.

By 2026, the AlphaFold database contains predicted structures for over 200 million proteins. The structural biology community has effectively solved the data scarcity problem it lived with for a generation. That fact alone warrants reflection before asking what pharmaceutical value it has created.

What Changed Immediately

The most direct impact was academic. Within a year of AlphaFold’s release, thousands of papers used its predictions to understand proteins with no known experimental structure — proteins implicated in disease that couldn’t be crystallized, expressed at scale, or solved by cryo-electron microscopy. Researchers working on neglected tropical diseases, rare genetic disorders, and understudied cancer drivers suddenly had structural hypotheses where they’d had none.

Related reading

This was genuine and important. A structural hypothesis is not a drug. But it is the beginning of a rational design process, and for the researchers working on targets where no structure existed, having one — even a predicted one — changed what experiments were worth running. Several programs that would have been stuck at the target biology stage, waiting for structural data that might take years to obtain experimentally, moved forward.

In drug discovery specifically, AlphaFold’s most immediate value was in hit validation and early lead optimization. Knowing the three-dimensional shape of a binding pocket, even approximately, lets medicinal chemists make better decisions about what chemical modifications will improve potency. Programs that used AlphaFold predictions in lead optimization advanced faster than historical comparators. The causal link is hard to isolate precisely — many things changed in those programs simultaneously — but the pattern is consistent enough to be suggestive.

AlphaFold 3, released in mid-2024, added meaningful capabilities beyond single-protein structure prediction. The new model predicts protein-ligand interactions (how drugs dock into binding sites), protein-nucleic acid complexes, and protein-protein interfaces with substantially improved accuracy over its predecessor. For medicinal chemists trying to understand how a drug candidate interacts with its target, this is a direct improvement in the tool’s practical utility.

The Caveats the Press Missed

The narrative that AlphaFold “solved drug discovery” was always wrong, and it was wrong in ways that took a few years to fully articulate. The most important limitation: AlphaFold predicts static structures — the lowest-energy conformation of a protein in idealized conditions. Most drug targets are not static. Proteins breathe, flex, adopt multiple conformations depending on their environment, binding partners, and post-translational modifications. A drug’s binding affinity often depends on catching a protein in a specific transient state that a static structure cannot capture.

This is why companies like Relay Therapeutics, long before AlphaFold, built their entire approach around protein dynamics. AlphaFold gave the field better starting points for structural hypothesis generation. It didn’t solve conformational complexity. A drug designed against a static AlphaFold structure might bind poorly if the relevant binding pocket only opens in certain dynamic states — states that require molecular dynamics simulation and experimental validation to characterize.

The predictions are also more reliable for well-conserved protein families — kinases, proteases, structural domains that evolution has constrained across millions of years — and less reliable for intrinsically disordered proteins, which constitute a substantial fraction of therapeutically interesting targets. Intrinsically disordered proteins (IDPs) don’t have a stable structure by definition; they exist as ensembles of conformations that change depending on context. AlphaFold’s predictions for them carry wide uncertainty bands that practitioners have to be careful not to treat as ground truth. AlphaFold 3 improved IDP handling and added better uncertainty quantification in these regions, but the underlying biological reality means that no static structure model will ever fully characterize these proteins.

Perhaps most importantly: structure prediction is not affinity prediction. Knowing where a drug might bind is not the same as knowing how strongly it will bind, how selective it will be for one target over others, or whether binding will produce the intended biological effect. The path from structure to therapeutic effect runs through experimental biology that no prediction model can shortcut. This seems obvious when stated directly. It was apparently not obvious enough to prevent a wave of premature optimism about AlphaFold’s pharmaceutical impact in 2021-2022.

The Cryptic Pocket Problem

Related reading

One area where AlphaFold has generated genuine pharmaceutical excitement is cryptic binding sites — pockets that only appear in certain protein conformations and were invisible to static crystal structures. Several programs now use AI-driven molecular dynamics simulation, seeded with AlphaFold structures, to explore conformational space and identify these transient pockets.

This matters because the pharmaceutical industry has spent decades lamenting the “undruggable proteome” — the large fraction of disease-relevant proteins with no obvious binding site for a small molecule. Standard drugging approaches require a pocket: a defined cavity where a small molecule can bind with sufficient surface area to achieve useful affinity. Many proteins don’t have obvious pockets in their most stable conformation but do in transient states that are accessible during normal biological function. If cryptic sites expand what’s druggable, the addressable market for small molecule drugs grows substantially.

There are now published examples of cryptic pocket identification leading to viable chemical series. Several academic groups and companies including D.E. Shaw Research have published cryptic pocket studies across oncology targets. None has yet reached Phase II trials, but the chemical matter is real and advancing. The field is watching carefully because the ability to drug previously inaccessible surfaces would represent a genuine expansion of what’s therapeutically tractable, not just a speedup on what was already possible.

The Manufacturing Gap

The less-discussed consequence of AlphaFold’s protein structure database is what it enabled in biologics manufacturing. Understanding protein structures at scale has accelerated antibody engineering, enzyme optimization, and protein stability engineering for biomanufacturing. Novo Nordisk and Eli Lilly — both producing GLP-1 receptor agonists at quantities the pharmaceutical industry had never previously imagined — have applied AI structural tools to improve the stability and manufacturability of their formulations. This is unsexy compared to “AI discovers new drug,” but the economic impact on the supply chains of the world’s most commercially successful drugs is substantial and ongoing.

Similar applications have emerged in industrial biotechnology, where AlphaFold-assisted enzyme engineering has improved the economics of bio-based manufacturing for chemicals and materials. Ginkgo Bioworks, Zymergen (before its acquisition), and several synthetic biology companies have incorporated AlphaFold-based protein engineering into their core workflows. This has nothing to do with human therapeutics, but it represents the database’s broader value — it became shared infrastructure for biology in the same way that GPS became shared infrastructure for navigation.

Academic vs. Industrial Impact

There’s a useful distinction between AlphaFold’s impact on academic research and its impact on drug development timelines and success rates. In academia, the impact has been immediate, broad, and clearly positive. Structural biology papers that previously required years of experimental work can now start from a reasonable computational hypothesis. Entire research programs that were impossible due to structural data scarcity are now feasible, particularly in neglected disease areas where no company has a financial incentive to fund expensive structural work.

In industrial drug development, the impact has been real but more concentrated. The bottleneck in drug development was never primarily structure determination. Even before AlphaFold, large pharma companies had structural programs that could solve crystal structures of their key targets on reasonable timescales. AlphaFold removes a cost and timeline component from that process, but it doesn’t address the bigger bottlenecks: clinical attrition, translational failure from animal models to humans, patient recruitment for trials, regulatory timelines. A faster path to the starting line doesn’t shorten the race.

The companies best positioned to extract value from AlphaFold are those working on structurally novel targets — proteins that had no experimental structure and were therefore genuinely inaccessible to structure-based drug design before 2021. A few startup biotechs have built programs around exactly these targets, and their early-stage pipelines are meaningfully differentiated from what was possible five years ago. Whether those programs survive to clinical proof-of-concept is the open question.

Related reading

What 2026 Actually Shows

By mid-2026, AlphaFold’s pharmaceutical impact is real but not yet transformative at the level of clinical outcomes. No drug has been approved where AlphaFold was the essential enabling technology. Several Phase II programs rely on AlphaFold-derived structural insights, and if any of them succeed, the causal narrative will be credible if not fully provable.

What has changed, durably and irreversibly, is the baseline of structural knowledge available to every researcher working on every disease. That’s not a drug. It’s something more like a new scientific instrument — one that takes years to generate its full impact because the impact comes from the experiments it enables, not from the instrument itself. The X-ray crystallography revolution of the mid-20th century, which gave structural biology its first systematic access to molecular structures, took roughly two decades to produce its first directly structure-enabled approved drug. AlphaFold’s revolution is faster in diffusion but probably not faster in the biology it enables.

Anfinsen’s observation in 1972 was that structure follows sequence. AlphaFold proved we can predict that transition computationally. The next problem — and it remains deeply unsolved — is predicting what drugs will do to living systems once they bind to the structures we now know. That’s a different class of question, requiring different data, different models, and a kind of wet biological knowledge that no structure prediction system can generate on its own.