Why AlphaFold Didn't Solve Protein Folding — Pushmeet Kohli, Google DeepMind & Sal Candido, Biohub

Summary · 4 min read instead of 33 min

Pushmeet Kohli is at Google DeepMind, where he worked on AlphaFold alongside John Jumper; he describes a path from security research through computer vision and deep learning into science. Sal Candido is at Biohub, whose team trains protein language models and works in the open with the scientific community toward Biohub's mission to cure all disease.

Kohli says AlphaFold did not solve protein folding: it learned to reproduce structures deposited in the PDB, while protein dynamics, disorder and context-dependent shape remain open. Both argue that scaling only pays off once the right data and the right scaling law are found, and Kohli reads Rich Sutton's bitter lesson as a warning to put the problem before any favored method. They weigh handcrafted inductive bias against general architectures, with Candido saying scaling itself takes real craft. On interpretability, Candido says protein language models hold unexploited knowledge of function and motion, while Kohli says calibrated confidence and behavioral characterization matter more than human-readable mechanisms. On the clinic, Kohli says AI is already used across drug discovery, and Candido urges aiming for 10x rather than 10% improvements.

Minute by minute

1. 1:39

   Scaling laws must be found first. Sal Candido says a common misconception is that scaling laws always exist; much of the work is finding a setup where more compute and data actually yield better results. Once that happens you can "turn the crank" and it becomes engineering. That depends on architecture and on data that carries the information needed for the problem.

2. 2:52

   Low-quality data can still help. Candido admits he is "lazy" and works with available data. Biohub trains protein language models on metagenomic sequences, many of which are not even whole proteins, yet they improve design of real proteins. The danger is scaling only easy-to-generate data, so he argues modelers and data generators should work together openly to identify the data actually needed.

3. 5:30

   The bitter lesson means problem first. Pushmeet Kohli, who was at DeepMind when Rich Sutton lectured there, reads the bitter lesson as a warning against approaching problems in a religious way as a modeler or a data person. The problem comes first and the solution space should stay flexible. He criticizes machine learning researchers who treated fixed train and test sets as given and only optimized models.

4. 7:41

   AlphaFold bet on modeling, virtual cell lacks data. For AlphaFold, Kohli says DeepMind lacked the expertise and resources to expand the PDB, so it invested in modeling. In cell genomics the team concluded the data was not there yet to build a virtual cell. His advice to newcomers is to be multidisciplinary, understand the problem and constraints such as data or model size, and fail fast.

5. 10:37

   Inductive bias versus general architectures. Kohli says AlphaFold 2's handcrafted design baked in biophysical intuition about residues influencing each other, making it more data efficient, and that curating good data is an art in itself. Candido says inductive bias helps with small data but can hold models back as data grows. He adds that scaling involves heavy algorithmic craft, with architectures being modified beyond standard transformers to fit their purpose.

6. 15:16

   Protein folding is far from solved. Kohli says "Proteins are not blocks": they are complex, disordered and change shape by context. He and John Jumper discussed that AlphaFold does not capture the true ground state or distribution of structures; it replicates structures someone deposited in the PDB, which happens to be useful. He urges continued funding for structure prediction and protein dynamics.

7. 18:21

   Cryo-EM micrographs as untapped source. Asked what he would wish into existence, Kohli says the computer vision researcher in him wanted to work directly on cryo-EM micrographs rather than processed PDB structures. Models that scale on that data could extract dynamics and distributional information. He tried it and says it needs more work.

8. 19:40

   Move from spokes to whole bicycles. Candido compares current models to modeling a single spoke when people want to understand a whole bicycle. He argues models must move from individual proteins to wheels and whole systems, placing them in biological context to learn broader interactions. He still supports continued work on folding models.

9. 21:16

   Protein language models hide unlocked knowledge. Responding to Feynman's quote on creating and understanding, Candido notes people designed with black boxes long before AI, but values understanding. Biohub's interpretability work finds protein language models encode not only structure but information about functions and motions. He says much remains to be unlocked from existing models, which compress evolutionary information.

10. 24:30

    Calibrated confidence earns trust. Kohli says AlphaFold 2 scored 90 GDT, but even at 95 GDT nobody would trust it if its pLDDT confidence were uncalibrated, since users could waste a year on a confidently wrong answer. DeepMind made progress understanding how AlphaFold 2 behaves. After the weights were released, people found it was a good disordered protein predictor, showing generalization the team had not discovered.

11. 26:27

    Interpretability depends on who interprets. Kohli says AlphaFold 2 is not interpretable to a human brain with its cognitive limits, but might be to a larger model. He suggests future frontier models given access to activation layers might predict AlphaFold's outputs and produce a theory of how it works. For users, behavioral characterization of strengths and limits is essential.

12. 28:26

    AI in the clinic and 10x thinking. Kohli calls the question of AI reaching the clinic ill-posed, since AI is already used across drug discovery; true 10x or 100x acceleration requires better biological models. Candido won't predict when a fully AI-made drug arrives but expects rapid progress. He says aiming for 10x rather than 10% opens broader solutions, and Biohub's mission to cure all disease demands that approach.

Published Oct 10, 2026.