AlphaFold solved a fifty-year-old problem and deserves every superlative it got. But the sentence "AlphaFold proves AI will solve science" does something quietly dishonest: it treats a spectacular exception as a general template. AlphaFold worked because protein structure prediction satisfied four conditions that most scientific problems do not, and the "AlphaFold for X" pitches that followed — for the cell, for materials, for disease efficacy, for whatever X you're raising money against — succeed or fail almost entirely on whether X inherits those conditions. None of this is an argument for AI-pessimism. It's a diagnostic you can run in an afternoon, before you spend three years and a Series B discovering which kind of problem you actually had.
Let me name the conditions precisely, because the vague version — "lots of data plus a big model" — is exactly what misleads.
The four conditions that made folding tractable
A large, high-quality, standardized labeled dataset. By the time AlphaFold2 trained, the Protein Data Bank held on the order of 200,000 experimentally-determined structures, deposited over five decades of X-ray crystallography, NMR, and cryo-EM. This is not "lots of data" in the web-scrape sense. It is curated, standardized, cross-validated data: shared file formats, resolution metadata, refinement statistics, agreed quality flags. Every structure was produced by mature methods whose error modes are understood. The corpus is large in the way that matters — large relative to the complexity of the function being learned, and clean enough that the labels point at the same thing regardless of which lab deposited them.
A crisply defined objective with a verifiable, quantitative success metric. The task is a function: amino-acid sequence in, 3D atomic coordinates out. There is no ambiguity about what a correct answer looks like, and there is a scalar that measures how close you are — GDT, the Global Distance Test, comparing predicted coordinates against a solved structure. Crucially, this was scored in CASP, the biennial blind competition John Moult started in 1994, where targets are structures that have been solved but not yet released. You cannot memorize or negotiate your way to a good CASP score. AlphaFold2's CASP14 result — a median GDT_TS around 92.4, where roughly 90 is considered competitive with experiment — was earned against ground truth the model had never seen and could not have leaked. The objective and the referee both existed before the model did.
An underlying regularity real enough to be learned. This is the condition people forget, and it is the deepest one. Christian Anfinsen won a Nobel in 1972 for the thermodynamic hypothesis: for many small globular proteins, the native fold is determined by the sequence — the sequence encodes the lowest-free-energy structure. That means a stable function f: sequence → structure genuinely exists to be approximated. AlphaFold does not simulate folding physics from scratch; its actual mechanism leans heavily on evolutionary signal, reading multiple sequence alignments to extract coevolutionary constraints — residues that mutate together tend to touch in 3D. But that only works because evolution has been running the folding experiment for billions of years and leaving the answer in the statistics. The regularity is real, it is stable across labs and days, and it left a legible trace. No such trace, no learnable function, no amount of compute helps.
Ground truth that is expensive but unambiguous. A crystal structure costs months and real money to obtain. But once obtained, it is not contested. Two structural biologists looking at the same well-resolved density map do not disagree about where the backbone goes. The label is costly at the margin and settled at the core. That combination — expensive enough to be scarce, unambiguous enough to be trusted — is what lets a held-out test set mean something.
Hold these four together and you can see why folding was tractable: a large standardized corpus, a crisp verifiable objective, a genuinely learnable regularity, and unambiguous ground truth. Score a problem 4/4 and a large model is not just applicable, it is close to the correct tool. AlphaFold is what 4/4 looks like.
Most of science is not 4/4
Now walk the same four conditions across the problems people actually mean when they say "solve science" — drug efficacy, disease mechanism, most of systems biology — and watch the score collapse.
The labels are scarce and expensive in a way the PDB is not. A representative efficacy or perturbation dataset might have 80 samples against 20,000 measured genes. Feature count exceeds sample count by two or three orders of magnitude, which means almost any label is perfectly separable by some feature combination; fitting the data is never the question, and the model will always find a separator. Compare that regime to 200,000 standardized structures and the asymmetry is the whole story.
The labels are also confounded and non-reproducible. This is the part I've argued at length in the ground-truth bottleneck in AI drug discovery, so I'll compress it: batch effects mean the model often learns "which Tuesday the sample was processed" rather than "sick versus healthy," and structured noise of that kind does not average out with more data — more capacity learns the artifact more precisely. Large replication efforts in cancer biology have found that more than half of landmark preclinical findings failed to reproduce the original effect. Sit with what that does to condition four. The ground truth is not merely noisy around a true value; a meaningful fraction of the labels point at effects that do not exist. Condition four inverts: the labels are expensive and contested.
Condition two fails just as hard, and it fails independently of data quality — a point garbage-in-garbage-out misses. "Does this molecule treat the disease" is not a crisp function with a verifiable metric the way "where does the backbone go" is. The outcome is stochastic, mediated by dozens of confounded systems, and measured years downstream in a Phase II readout that itself has wide error bars. There is no CASP for efficacy, no blind referee holding the answer before you predict. When the objective is ill-defined and the success metric is contested, a leaderboard number tells you almost nothing about whether you learned biology.
And condition three — a real learnable regularity — is often simply unknown. Anfinsen guaranteed that a folding function exists. Nothing guarantees that "sequence → clinical outcome" is a stable function at all, rather than a path-dependent, context-soaked process where the same input yields different outcomes in different bodies on different days.
Here is the arithmetic that makes the gap concrete. AlphaFold learned from roughly 2 × 10⁵ standardized, uncontested labels of a function that provably exists. A typical efficacy model learns from ~10¹ to 10² noisy, confounded, partly-irreproducible labels of a function that may not exist as a stable object — across a feature space of ~10⁴. Same model class, incomparable footing. The architecture was never the variable that moved.
MAMMAL versus AlphaFold3: watch a condition weaken in real time
You don't have to leave protein modeling to see the conditions fail. You just have to change the task slightly and watch confidence become a leaky proxy.
AlphaFold3 extends structure prediction to complexes, ligands, and nucleic acids, and it is a landmark in its own right. But consider a recent, careful comparison I've written up in full — why a sequence-only model out-predicted AlphaFold3 on antibody binding. On the specific task of separating antibody binders from non-binders across seven targets, a 458M-parameter sequence-only model (MAMMAL, from IBM Research) had higher AUROC than AlphaFold3 in five of seven cases — and on one target AF3 scored below chance.
The naive read is "sequence beat structure." The correct read is that the task changed and one condition quietly broke. AF3 was never trained to answer "does this bind"; it was run zero-shot, with its structural-confidence scores pressed into service as a proxy for binding likelihood. A confidence-in-structure score is not a binding label. It is a proxy, and proxies stay coupled to the truth only where the conditions that made them trustworthy still hold — rigid, cooperative, crystallizable targets like the globular TBG, where AF3 scored a perfect 1.00. On flexible, glycosylated, disorder-rich targets — the messy majority of therapeutically interesting proteins — the proxy decoupled and AF3 was confidently wrong. MAMMAL, by contrast, had been trained on the actual binds/doesn't-bind label, with negative examples.
That is the whole diagnostic in miniature. AF3 is a 4/4 model on the task it was built for. Push it onto a task where the objective is different and the ground truth (binding under real conditions) is no longer what the training signal measured, and you don't get a graceful degradation. You get high confidence pointed at the wrong answer. The failure mode of a strong model outside its conditions is not weakness. It is precision aimed at the wrong function.
The diagnostic: score before you pitch
So before you say "AlphaFold for my field," score your field. Rate each condition 0, 1, or 2, and read the total honestly.
| Condition | Score 2 (like folding) | Score 0 (unlike folding) |
|---|---|---|
| 1. Standardized labeled corpus | Large, curated, cross-validated, consistent across labs (PDB) | Scarce, per-lab, confounded, batch-effect-ridden |
| 2. Crisp objective + verifiable metric | Well-posed function, blind held-out referee (sequence→structure, CASP/GDT) | Contested outcome, no agreed metric, no blind test |
| 3. Real learnable regularity | Provably exists, stable, leaves a legible trace (Anfinsen + coevolution) | Unknown whether a stable function exists at all |
| 4. Ground truth expensive but unambiguous | Costly to get, uncontested once got (a crystal structure) | Costly and contested; partly irreproducible |
A 7 or 8 predicts real acceleration — and this is not a biology-only claim. Medium-range weather forecasting scores high on all four: decades of standardized reanalysis data, a crisp objective, stable atmospheric physics, and unambiguous verification against what actually happened. Learned models have posted genuine gains there, and the diagnostic predicts exactly that. AlphaFold for weather is real precisely because weather is 4/4. That is the optimistic half of the argument, and it is the half people skip: the conditions are a map of where to bet, not just where to worry.
A 3 or below predicts something specific and worse than failure. Not "the model won't fit" — it will fit beautifully, because with 10⁴ features and 10² samples everything fits. It predicts a model that is confident and wrong, delivered faster with more compute. That is the most expensive outcome in science, because a hedged wrong answer gets checked and a confident wrong answer gets funded.
The middle scores are where judgment lives. A 4 or 5 usually means one condition is doing all the damage, and naming which one tells you where to spend. If conditions 1, 2, and 3 are strong but 4 is weak — good learnable function, clean objective, but contested ground truth — your money goes to the wet lab, not the GPU cluster, because you are label-limited, not model-limited. If 4 is strong but 2 is weak — you can measure the answer unambiguously but nobody agrees what question you're answering — you have a specification problem, and no dataset fixes an ill-posed objective.
The seductive error is to look at AlphaFold, see "big model plus big data," and conclude the recipe transfers. It doesn't, because the recipe was never the model. It was the four conditions the model got to stand on. AlphaFold is what happens when a hard problem is secretly well-posed. Most hard problems are hard because they are not.
Run the score first. Where it comes back high, build the thing — the acceleration is real and it is coming. Where it comes back low, the honest move is to stop looking for a bigger model and go make the ground truth the model would need. AlphaFold didn't skip that step. Nature and fifty years of crystallographers had already done it.