From DENDRAL to AlphaFold: what let AI into science was the assessment

On 8 October 2024 the Nobel Prize in Physics went to John Hopfield and Geoffrey Hinton. The next day, half the Chemistry prize went to David Baker, and the other half jointly to Demis Hassabis and John Jumper^[11].

Two prizes, two days, both landing on machine learning.

Reading that as "AI finally got good enough" misses what actually happened. The prize was not given to a model. It was given for a problem that stopped being open — protein folding, on the page of every textbook since Anfinsen argued that sequence determines structure^[1]. And what made the answer acceptable to the field was not an architecture. It was an assessment regime that had been hiding the answer in advance since 1994.

This is that sixty years, walked backwards.

1965: writing the expert down as rules

The first serious attempt to put AI inside science was DENDRAL, begun at Stanford in 1965 by the geneticist Joshua Lederberg and the computer scientist Edward Feigenbaum. The task was narrow and real: given mass spectrometry data, infer the molecular structure of the compound^[2].

The method was to take how an organic chemist reasons about structure and write it down, rule by rule, then generate candidates with an algorithm that could enumerate every legal structure and let the rules cut them down. It worked. Within narrow compound classes its inferences were usable, and it gave the field the term "expert system".

Its ceiling was equally clear. The rules had to be written by a person, and the set of domains you can finish writing is small. Every new chemical family meant sitting another chemist down and getting them to say their intuitions out loud. The bottleneck was not compute. It was knowledge acquisition.

1994: a contest that hides the answer

The pivotal event of the second stage was not a model.

In 1994 Moult and colleagues ran the first CASP — Critical Assessment of Structure Prediction. The rules: the organisers collect proteins whose structures have been solved but not yet released, publish only the sequences, take predictions by a deadline, and reveal the answers afterwards^[3].

What that design fixed was a problem running through the whole field. Every group claimed its method worked, and every group had tested on examples it had chosen itself. CASP replaced "we say it works" with "does it work on problems you have not seen", every two years, ranked in public.

Whether a field can be moved by AI depends first on whether it can hide the answer in advance.

From the 1990s into the early 2010s, sequence alignment, hidden Markov models and coevolution analysis climbed the CASP rankings without any round approaching experimental accuracy. Two assets accumulated in that period: tens of thousands of experimental structures in the Protein Data Bank, and a scoring regime nobody could game.

After 2012: representation learning takes over

AlexNet's ImageNet result in 2012 retired the assumption that features must be designed by hand^[4]. What mattered for science was not image recognition. It was that the layer that previously had to be written down by an expert could now be learned from data. The wall DENDRAL hit was taken down from the other side, almost fifty years later.

DeepMind's first AlphaFold won CASP13 in 2018 by predicting distance distributions between residue pairs with a deep network, then using gradient descent to find a structure consistent with those distances^[5]. Better than the competition, still short of experimental accuracy.

2021: AlphaFold2, and why it was believed

At CASP14 in 2020, AlphaFold2 was, in its authors' own summary, the first computational method able to predict structures with atomic accuracy even where no similar structure was known — competitive with experimental structures in a majority of cases, and greatly outperforming every other method entered^[6]. Architecturally it trained multiple sequence alignments and a structural representation end to end in one differentiable model, and emitted a per-residue confidence estimate alongside the prediction.

That confidence estimate (pLDDT) is worth pausing on. It tells the user which regions the model is sure about and which it is not — the model hands over not just an answer but a list of its own suspect parts.

But what changed how structural biologists worked, within months, was the venue. The same number from a test set DeepMind had chosen would have been a paper. From CASP, it was a conclusion.

The AlphaFold Protein Structure Database then released predictions covering over 214 million sequences^[7], and AlphaFold 3 extended prediction from single proteins to complexes with nucleic acids, small molecules and ions^[10].

Baker's half: prediction inverted into design

The other half of the Chemistry prize is a separate line.

David Baker's lab had been developing Rosetta since the 1990s — energy functions and conformational search. In 2003 they published Top7: a protein with a fold not found in nature, designed computationally, expressed, and then solved by X-ray crystallography to match the design^[8]. That is a reversal of direction. Not guessing at a structure nature already has, but building one it does not.

RFdiffusion put a diffusion model on that line in 2023, generating backbones under constraints — what it must bind, what symmetry it must have — then designing sequences for them, with experimental validation^[9].

So the two halves of the 2024 Chemistry prize are two directions through the same map: reading sequence into structure, and running it backwards.

What the 2024 prize was actually for

Nobel prizes are not given to tools. The Chemistry citations read "computational protein design" and "protein structure prediction"^[11] — a change in the state of a problem, not a method.

Both halves satisfy the same three conditions:

Condition	How protein structure met it
Large, standardised, trustworthy prior data	Decades of experimental structures in the PDB
An assessment that hides the answer in advance	CASP, biennial, since 1994
A validation path that does not run through the model	X-ray, cryo-EM, actually making the protein

The third is the one most often skipped. An RFdiffusion design does not count because the model says it will fold that way. It counts once it has been made and measured.

The same conditions, in other fields

Those three conditions explain which fields moved quickly.

Materials: GNoME trained graph networks at scale to search for inorganic crystals and reported 2.2 million structures below the current convex hull — and, to the point of the third condition, that 736 of the stable ones had already been independently realised experimentally^[12]. Weather: GraphCast produced medium-range global forecasts with a graph neural network, outperforming the most accurate operational deterministic systems on 90% of 1,380 verification targets^[13]. Both had large prior datasets and established evaluation practice — weather especially clean, because tomorrow reveals the answer by itself.

For a wider survey, Wang and colleagues' Nature review organises this wave into data representation, hypothesis generation and experimental design^[16].

Where the prize does not reach

The same methods, in fields missing those conditions, produce results that look just as good and that nobody can refute.

Kapoor and Narayanan examined reproducibility after machine learning entered scientific research, identifying 329 papers across 17 fields affected by data leakage — information bleeding between training and test sets, so reported performance exceeds real performance^[14]. These errors share a property: they do not crash anything. They only make the number better.

Materials has a concrete counter-example too. Leeman and colleagues re-examined claims from high-throughput inorganic prediction and autonomous synthesis, and found that a substantial share of the reported "new materials" corresponded on closer comparison to known phases, or rested on structural determination too weak to support the claim^[15].

The difference from AlphaFold is not model quality. It is whether anything outside the model can reject the answer. CASP can. X-ray can. Tomorrow's weather can. A paper reporting AUC on its own split cannot.

What this review leaves out

A review is not an evaluation. Each stage above is represented by one or two pieces of work and skips a great deal that ran alongside it — the whole coevolution-analysis lineage, design methods other than Rosetta, assessments other than CASP such as CAMEO and CAPRI. The reference list is an entry point, not a map.

Every figure quoted here comes from its cited source, and every reference below was checked against the metadata registered for its DOI. None of the numbers are ours. Two that circulate widely were cut during that check because they could not be traced to the papers cited here — a CASP14 median score, and a count of stable materials — which is worth knowing if you have seen this history told with them in.

Next

The next part goes into the literature itself. When a question has no CASP, how do you build an extraction table from two hundred papers that you would actually run statistics on — provenance per field, the failure modes of extraction, and which errors code can catch versus which need a person to read.

References

1. [1]Anfinsen, C. B. (1973). Principles that govern the folding of protein chains. Science 181(4096), 223–230. doi:10.1126/science.181.4096.223
2. [2]Lindsay, R. K., Buchanan, B. G., Feigenbaum, E. A. & Lederberg, J. (1993). DENDRAL: a case study of the first expert system for scientific hypothesis formation. Artificial Intelligence 61(2), 209–261. doi:10.1016/0004-3702(93)90068-M
3. [3]Moult, J., Pedersen, J. T., Judson, R. & Fidelis, K. (1995). A large-scale experiment to assess protein structure prediction methods. Proteins: Structure, Function, and Bioinformatics 23(3). doi:10.1002/prot.340230303
4. [4]Krizhevsky, A., Sutskever, I. & Hinton, G. E. (2017). ImageNet classification with deep convolutional neural networks. Communications of the ACM 60(6), 84–90. doi:10.1145/3065386
5. [5]Senior, A. W., Evans, R., Jumper, J. et al. (2020). Improved protein structure prediction using potentials from deep learning. Nature 577, 706–710. doi:10.1038/s41586-019-1923-7
6. [6]Jumper, J., Evans, R., Pritzel, A. et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature 596, 583–589. doi:10.1038/s41586-021-03819-2
7. [7]Varadi, M., Bertoni, D., Magana, P. et al. (2024). AlphaFold Protein Structure Database in 2024: providing structure coverage for over 214 million protein sequences. Nucleic Acids Research 52(D1), D368–D375. doi:10.1093/nar/gkad1011
8. [8]Kuhlman, B., Dantas, G., Ireton, G. C. et al. (2003). Design of a novel globular protein fold with atomic-level accuracy. Science 302(5649), 1364–1368. doi:10.1126/science.1089427
9. [9]Watson, J. L., Juergens, D., Bennett, N. R. et al. (2023). De novo design of protein structure and function with RFdiffusion. Nature 620, 1089–1100. doi:10.1038/s41586-023-06415-8
10. [10]Abramson, J., Adler, J., Dunger, J. et al. (2024). Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500. doi:10.1038/s41586-024-07487-w
11. [11]The Royal Swedish Academy of Sciences (2024). The Nobel Prize in Chemistry 2024 — press release. Nobel Prize Outreach. https://www.nobelprize.org/prizes/chemistry/2024/press-release/
12. [12]Merchant, A., Batzner, S., Schoenholz, S. S. et al. (2023). Scaling deep learning for materials discovery. Nature 624, 80–85. doi:10.1038/s41586-023-06735-9
13. [13]Lam, R., Sanchez-Gonzalez, A., Willson, M. et al. (2023). Learning skillful medium-range global weather forecasting. Science 382(6677), 1416–1421. doi:10.1126/science.adi2336
14. [14]Kapoor, S. & Narayanan, A. (2023). Leakage and the reproducibility crisis in machine-learning-based science. Patterns 4(9), 100804. doi:10.1016/j.patter.2023.100804
15. [15]Leeman, J., Liu, Y., Stiles, J. et al. (2024). Challenges in high-throughput inorganic materials prediction and autonomous synthesis. PRX Energy 3(1), 011002. doi:10.1103/PRXEnergy.3.011002
16. [16]Wang, H., Fu, T., Du, Y. et al. (2023). Scientific discovery in the age of artificial intelligence. Nature 620, 47–60. doi:10.1038/s41586-023-06221-2