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

What it claims

This is a 30-minute conference panel, not a paper, so it reports no new results. It carries a few first-hand observations from the people who built these systems that the summary flattens:

• What AlphaFold 2 actually optimised. Kohli says the target was never the true ground state or the distribution of conformations a protein takes. It was to reproduce the single structure someone had deposited in the PDB. The "solved" framing applies only to that proxy. Dynamics, disorder and context-dependent shape remain open.
• A failed virtual-cell attempt. DeepMind applied its AlphaFold-style approach to cell-by-gene data and concluded "after a lot of work" that the data was not yet sufficient to build a virtual cell. This is a concrete negative result for anyone planning to model single-cell expression data.
• Inductive bias has a tipping point. Candido says handcrafted bias pays off at small data scales. At larger scales a bias that is "not exactly correct" can hold the model back. AF2's residue-interaction bias was justified because the PDB was fixed and could not be expanded by an order of magnitude.
• Noisy data can help. Training protein language models on metagenomic sequences, many of which are "not even a real, whole protein," still improved design of working proteins and understanding of known ones. Kohli adds that replicating the same data gains nothing, so coverage matters more than volume.
• Protein language model representations hold more than structure. Biohub's interpretability work finds information about function and motion in the representations, not only structure.
• Trust comes from calibration, not mechanism. Kohli puts AF2 at a GDT of about 90 on the evaluation set of the time. He argues a GDT of 95 with uncalibrated pLDDT would have been untrustworthy. AF2's usefulness as a disorder predictor was found by outside users only after the weights were released, not characterised before launch.

Limits

• Every claim is an opinion or anecdote from a panel. There are no ablations, datasets, sizes or compute figures.
• The GDT ≈ 90 figure comes with no named benchmark beyond "that set at the time," and no baseline.
• Kohli's attempt to train directly on cryo-EM micrographs is reported only as "tried it, but it requires more work," with no details.
• The cell-by-gene result gives no detail on which data or models were tried.
• The metagenomic-data claim names no model, dataset or metric.
• The panel gives no timeline for AI-driven drug discovery. Both speakers decline to quantify one.
• Read it for framing and for the negative results above, not for anything you can reproduce or adopt.