What AlphaFold Actually Delivered: A Drug Pipeline Reality Check

The announcement landed in late 2020 with the kind of clean narrative that science rarely produces: DeepMind had solved protein folding, a fifty-year-old grand challenge of biology. The coverage was appropriate. The problem was genuinely hard, and the solution was genuinely elegant. What happened next is more complicated, and more interesting, than most of the anniversary coverage will admit.

By 2029, AlphaFold and its successors — RoseTTAFold, ESMFold, Chai-1, the half-dozen models that followed — have produced around 400 million predicted structures. That number is enormous. The question worth asking is what we have done with those structures, because the gap between structural knowledge and therapeutic action turns out to be substantial, and closing it has required a different kind of work than anyone quite expected.

Start with the wins, because there are real ones. The clearest contribution has been in what the industry calls target identification — the earliest stage of drug discovery, where you’re trying to understand which proteins are even worth pursuing. Before AlphaFold, the standard move was to work backward from known structures or to accept that large stretches of the proteome were simply inaccessible. Membrane proteins, intrinsically disordered regions, complexes that resist crystallization: these were the dark matter of structural biology. Predicted structures don’t fully resolve the disordered-region problem (disorder is a feature, not a bug — the structure is the function, and it’s dynamic), but they have opened up families of targets that pharma was previously ignoring on purely practical grounds.

The most documented success story involves the kinase-adjacent class of proteins. Kinases themselves were already well-characterized. The neighboring regulatory proteins — the phosphatases, the scaffolding proteins that hold signaling complexes together — were poorly understood structurally because they’d resisted the standard crystallography workflow. Several of the most promising oncology programs currently in Phase II trials are pursuing targets that were essentially opaque before 2021. That’s a real contribution.

Where the story gets more complicated is in the translation from structure to drug. The promise, stated in its aggressive form, was that knowing the structure of a protein would dramatically accelerate finding small molecules that bind to it. The reality is that structure is necessary but nowhere near sufficient. A protein’s shape tells you roughly where a small molecule might fit; it doesn’t tell you whether binding there will have any therapeutic effect, whether the molecule can reach the target inside a cell, or whether the cell will develop resistance mechanisms in three months.

The drug discovery workflow that has actually emerged is more of a collaboration than a handoff. AlphaFold provides a structure. Molecular dynamics simulations — themselves accelerated by AI, using models like the various successor architectures to OpenFold — identify the dynamic pockets, the regions that open and close on timescales of microseconds. Generative chemistry models then propose molecules against those dynamic pockets. Human medicinal chemists evaluate the proposals, throw out around 80 percent of them for reasons that are hard to formalize (synthetic accessibility, metabolic stability patterns they recognize from experience, off-target liability concerns), and iterate. The human judgment hasn’t been removed from the loop; it’s been repositioned earlier in the process, filtering at the hypothesis stage rather than the synthesis stage.

This matters for understanding the timelines. The early projections — some serious scientists said this, not just journalists — suggested that AlphaFold might compress drug discovery timelines from 10-15 years to 3-5 years. The honest 2029 assessment is that it has compressed specific stages of target identification and lead generation by meaningful amounts, but hasn’t touched the clinical development timeline, which is dominated by biology and regulatory review rather than structural knowledge. Phase III trials still take as long as they take. The immune system’s response to a candidate drug doesn’t care how good your structure prediction was.

The more profound shift, and one that gets less press coverage, is in academic biology. The 200 million structures in the AlphaFold Protein Structure Database have changed what questions structural biologists ask in the first place. When every structure is known (or predicted with reasonable confidence), the interesting questions are no longer “what does this protein look like?” but “how do these proteins find each other, how do they change shape in response to signals, and how do those changes cascade through cellular networks?” The field has moved toward dynamics and interaction, which is both harder and more biologically relevant.

A word on the failures, because they’re instructive. There was a substantial period in 2024 and 2025 where several biotech startups tried to use AlphaFold structures as essentially the only input into AI-driven drug design pipelines. Skip the medicinal chemistry iteration, go directly from structure to synthesizable candidate. The results were roughly what the skeptics predicted: high structural plausibility, low biological relevance. The molecules bound to the purified protein beautifully in vitro and did approximately nothing in cell-based assays. The lesson wasn’t that AI drug discovery doesn’t work; it was that the chain from predicted structure to clinical candidate has more links than the 2021 hype implied.

The companies that have actually advanced candidates to the clinic — and there are now several, with two having received regulatory approval — did so by treating AI as a tool that accelerates human decisions rather than replaces them. That sounds obvious in 2029. It was not obvious, and not unchallenged, in 2022.

One development that deserves more attention is what’s happened in tropical disease research, which has historically been starved of structural biology resources because the commercial return doesn’t support the infrastructure investment. The combination of cheap, fast structure prediction and open databases has allowed small academic groups in countries that couldn’t maintain protein crystallography facilities to do competitive structural biology. Several important antiparasitic leads have emerged from exactly this kind of distributed work. AlphaFold’s most lasting contribution may turn out to be democratization — not cutting years off blockbuster drug timelines, but giving everyone access to a capability that was previously reserved for well-funded institutions in wealthy countries.

The protein folding problem is solved. The drug discovery problem is not, and never was going to be, the same problem. What AlphaFold actually delivered was a complete, accessible map of protein structures — and a clearer view of how much terrain remains between that map and a medicine that helps a patient. That’s a genuinely significant contribution. It was always a narrower one than the initial coverage suggested.