The protein folding problem had a specific, brutal shape for half a century. A protein is a chain of amino acids — a string of beads, in the usual telling — and the instant a cell builds one, it collapses into an intricate three-dimensional knot. That final shape decides everything the protein does: whether it grabs a target molecule, flips a genetic switch, or ferries oxygen through your blood. Predicting the knot from the bead order alone was the puzzle, and for decades the chemistry was understood while the computation stayed hopeless.
Then it wasn’t. The story you have probably heard ends there — machine learning walks in, a 50-year problem falls, biology enters a new era. All of that is true. But it is the photograph of the achievement, and the achievement, fittingly, is better understood as a film. What AlphaFold solved was real and enormous. What it left untouched is the part that governs how proteins actually function — and the field has spent the years since quietly rebuilding it.
What did AlphaFold actually solve?
In November 2020, at the biennial CASP competition — a blind test where research groups predict structures that have been solved in the lab but not yet made public — DeepMind’s AlphaFold2 posted numbers the field had not expected to see for a decade. Its predictions reached a median GDT_TS of 92.4 (a score where 100 is a perfect match to the experimental structure, and anything above roughly 90 sits inside experimental error itself). The 2021 Nature paper reported a median backbone accuracy of 0.96 Å — under one angstrom, about the width of two atoms. By 2022, DeepMind and the European Bioinformatics Institute had released predicted structures for more than 200 million proteins, covering nearly every sequence in the public databases.
To feel the size of that, look at what it replaced. Determining a single protein structure by X-ray crystallography, cryo-electron microscopy, or NMR can take months to years of exacting lab work, and after decades of effort researchers had solved only around 100,000 unique structures — against billions of known protein sequences. AlphaFold closed most of that gap in about a year, and within two years more than a million researchers had drawn on the database. The bottleneck did not narrow. It effectively disappeared.
Now read the Nature paper’s own sentence carefully. It describes the result as solving “the structure prediction component of the ‘protein folding problem.’” That word — component — is the entire essay. AlphaFold’s authors were precise about the boundary of what they had done. The press release writers were not, and the gap between those two sentences is where most of the confusion about AI in biology now lives.
The full problem is really two questions. First: given a sequence, what stable shape does it settle into? Second: how does it move once it has settled — what other shapes does it visit, and how often? AlphaFold answered the first with near-experimental accuracy. The second it did not attempt, and was never built to.
Why isn’t a correct structure enough?
Here is the part the celebration skipped, and it is the mechanism beneath the mechanism. A protein is not its lowest-energy shape. It is the full set of shapes it can adopt and the rates at which it flickers between them — what structural biologists call its conformational ensemble. A predicted structure is a photograph. Function is a film.
That is not a metaphor reaching for effect; it is mechanically what is happening. A receptor on a cell surface does its job by flexing between an “off” shape and an “on” shape when a signal arrives. An enzyme yawns open, clamps onto its target, snaps shut, and releases. A single static structure — even a flawless one — is one frame lifted out of that motion. It tells you where the protein rests. It is largely silent about how it works. (This is established biophysics, not a fringe view — the role of conformational dynamics in function predates AlphaFold by decades.)
A predicted structure is a photograph. Function is a film. The Architect · June 2026
Take hemoglobin, the protein that carries oxygen in your blood. It works precisely because it changes shape: binding one oxygen molecule nudges the whole structure into a new conformation that grabs the next three more eagerly. That cooperative shift — not any single frozen pose — is the function. A perfect snapshot of either the oxygen-loaded or the empty state would miss the mechanism entirely, because the mechanism is the transition between them. Most of the machinery of life runs on exactly this kind of motion.
Why should you care about a frame rate? Because the places drugs bind, the shape changes that drive disease, and the specific motion a molecule has to interrupt all live in the film, not the photograph. You can get the resting shape exactly right and remain mute about the thing biology actually needed to know. Gregory Bowman, writing in the Annual Review of Biomedical Data Science in 2024, put the stakes plainly: calling structure prediction — let alone protein folding — “solved” is dangerous, because believing it could stall the work that still matters.
Where does the single-structure picture break down?
AlphaFold was trained, in effect, to map one sequence to one structure — to find the single most probable resting shape. That assumption holds for most proteins most of the time, which is exactly why the tool is so useful. It breaks in named, predictable places, and the places are worth knowing because they are where biology gets interesting.
Fold-switching proteins are the cleanest case. These adopt two entirely different stable shapes from one sequence — a molecular Necker cube. Standard AlphaFold returns one of them and stamps it with high confidence, giving no hint that a second exists. A 2024 analysis led by Lauren Porter’s group at the NIH found that when the model does surface an alternative fold, it often appears to be recalling a similar structure from its training data rather than reasoning about the protein’s energy landscape (theorized — the precise mechanism is still debated, but the failure is reproducible).
Then there are cryptic pockets: cavities that do not exist in the resting structure but open transiently as the protein breathes, and which are sometimes the only druggable foothold on an otherwise smooth surface. TEM β-lactamase — the enzyme bacteria use to chew up penicillin — has well-characterized cryptic pockets that appear only in higher-energy states. A photograph of its ground state shows a flat wall where, in motion, a door exists.
A starker case is the intrinsically disordered proteins — a large slice of the human proteome that never folds into one stable shape at all, instead sampling a shifting cloud of conformations as its job. For these, the single-structure question is not merely hard to answer; it is the wrong question. AlphaFold will still hand back a confident-looking structure, usually with low-confidence flags scattered across the disordered stretches, but the object it is trying to predict does not exist in the form the model assumes.
The commercially loudest example is GPCRs — G-protein-coupled receptors, the target of roughly a third of all approved drugs. They signal by switching between active and inactive shapes, and capturing the right one is the whole game in drug design. A further wrinkle, documented in The Protein Journal in December 2025, is that AlphaFold’s training data skews toward soluble, well-folded proteins, so membrane proteins and weak, transient interfaces — a large slice of drug targets — are precisely where it is least sure-footed.
Did it deliver for drug discovery?
Here honesty has to cut both ways, because the anti-hype case becomes its own distortion the moment it tips into anti-science. AlphaFold delivered real, measurable downstream wins, and the most rigorous critiques come from people who use it daily and want it to be better. In May 2024, DeepMind released AlphaFold3, a redesigned system that predicts not just single proteins but their complexes with DNA, RNA, small molecules, and ions — the interactions that actually constitute drug binding.
The wins are concrete and worth naming. AlphaFold has accelerated virtual screening, letting chemists triage enormous compound libraries against a predicted pocket before touching a pipette. It has helped researchers fit and interpret noisy cryo-EM density maps, turning ambiguous blobs of signal into models in hours rather than weeks. And it has fed structure-guided design across thousands of labs that could never have afforded the experimental structures they now start from. None of that is hype; it is the new baseline.
A 2025 benchmark of AlphaFold3 across drug-discovery tasks, posted to bioRxiv in April, mapped both edges of the tool with unusual care. On static protein-ligand interactions — a drug locking into a relatively rigid pocket — AF3 excelled, beating traditional docking software on side-chain orientation accuracy. For that class of problem it is genuinely, immediately useful, and that should not get lost in the caveats.
But the same benchmark found the failures, and they trace one fault line. AF3 struggled once binding involved a conformational change beyond about 5 Å. It showed a persistent bias toward the active shape of GPCRs even when modeling antagonist-bound states that should sit in the inactive shape. It performed poorly on ternary complexes — three molecules bridged together — which is exactly the geometry that PROTACs and molecular glues, a fast-growing class of degrader drugs, depend on. And it could not reliably rank compounds by binding strength. Every one of those is a dynamics problem wearing a different costume.
The scale of the bet riding on this is hard to overstate. Isomorphic Labs, the DeepMind spinout that co-developed AlphaFold3, raised $600 million in March 2025 and holds drug-discovery partnerships with Eli Lilly and Novartis worth a potential $3 billion in milestone payments. Tellingly, the company is not wagering that prediction is already enough — it is building an entire engine to push past the static structure toward the interactions and motions AlphaFold alone misses. The commercial frontier and the scientific frontier turn out to be the same line.
What is the field building now?
The useful way to read all of this is not as AlphaFold’s failure but as the next problem snapping into focus. Since 2020, structural biologists have been rebuilding the film one frame at a time, and the early moves were almost comically low-tech.
The first trick was to feed the model fewer evolutionary relatives. AlphaFold reads a stack of related sequences — a multiple sequence alignment — to infer structure; thin that stack out, and you destabilize the model’s certainty enough to coax alternative shapes out of it. Methods built on this idea, such as AF-Cluster (Wayment-Steele and colleagues, Nature, 2024), recovered multiple experimentally real conformations from a tool that was never designed to produce them. AlphaFold3’s diffusion-based architecture, which generates structures by denoising rather than direct prediction, can in principle sample a spread of shapes rather than commit to one — though whether that spread faithfully matches a protein’s true ensemble is, for now, an open question (theorized). Further out sit physics-ML hybrids that stitch AlphaFold’s shapes to molecular-dynamics simulation, aiming to recover not just the shapes but the motion between them (speculative, and an active research front).
What would success look like? Not a single sharper photograph but a model that hands you the whole reel — the populations of shapes a protein occupies and the odds of catching it in each. That is a harder problem than the one AlphaFold solved, and it is the one that decides whether a drug works.
Prediction is not understanding
So: did AlphaFold solve protein folding? It solved the component that had been the bottleneck for fifty years, and predicting a near-perfect static structure from sequence alone is not a small thing — it is one of the genuine scientific achievements of the century. But solving the photograph is what finally revealed how much of biology was always in the film: the motion, the alternative shapes, the transient pockets, the switching that turns a molecule from a sculpture into a machine.
The transferable idea is worth carrying past proteins. A model can be exactly right about the shape of a thing and completely silent about its life. The next time a system is announced as having solved a field, the sharpest question is the one DeepMind’s own paper asked and the headlines quietly dropped — which component?
Sources
Jumper, J. et al. (2021). “Highly accurate protein structure prediction with AlphaFold.” Nature 596:583–589. Primary source for the CASP14 result, the median backbone accuracy of 0.96 Å, and the paper’s own “structure prediction component of the protein folding problem” framing.
Abramson, J. et al. (2024). “Accurate structure prediction of biomolecular interactions with AlphaFold 3.” Nature 630:493–500 (May 8, 2024). Primary source for AlphaFold3’s diffusion architecture and extension to complexes, nucleic acids, and ligands.
Bowman, G. R. (2024). “AlphaFold and Protein Folding: Not Dead Yet! The Frontier Is Conformational Ensembles.” Annual Review of Biomedical Data Science 7:51–57. The dynamics critique from inside the field, including the cryptic-pocket example and the warning against calling the problem “solved.”
Zheng, H. et al. (2025). “AlphaFold3 in Drug Discovery: A Comprehensive Assessment of Capabilities, Limitations, and Applications.” bioRxiv (posted April 8, 2025). The drug-discovery scorecard: strong on static interactions and side-chain accuracy, weak past ~5 Å conformational change, biased toward active GPCR states, poor on PROTAC ternary complexes, unreliable at affinity ranking.
Wayment-Steele, H. K. et al. (2024). “Predicting multiple conformations via sequence clustering and AlphaFold2.” Nature 625:832–839. Representative of the MSA-subsampling methods (AF-Cluster) used to recover alternative conformations.
Chakravarty, D. et al. (2024). “AlphaFold predictions of fold-switched conformations are driven by structure memorization.” Nature Communications 15 (Porter group, NIH). Source for the fold-switching limitation and the structure-memorization finding.
Varadi, M. et al. (2024). “AlphaFold Protein Structure Database in 2024.” Nucleic Acids Research 52:D368–D375. Source for the scale of the public structure release (200M+ predictions, 214M sequences).
“Advantages and Limitations of AlphaFold in Structural Biology.” The Protein Journal (December 2025). Source for the training-data bias toward soluble, well-folded proteins and weaker performance on membrane proteins and transient interfaces.
Isomorphic Labs (March 31, 2025). “Isomorphic Labs announces $600 million funding to further develop its next-generation AI drug design engine.” Company announcement, with reporting from TechCrunch and CNBC on the Eli Lilly and Novartis partnerships (potential ~$3 billion in milestone payments). Used as the commercial-stakes anchor.