Why AlphaFold Didn't Solve Protein Folding — Pushmeet Kohli, Google DeepMind & Sal Candido, Biohub
- ID
- 33893
- Status
- summarized
- Published
- 10 Oct 2026, 8:31 AM
- Fetched
- 10 Oct 2026, 8:40 AM
- Provider
- Latent Space
- Category
- developer-ai
- Original URL
- https://www.latent.space/p/biohub-deepmind
- Source URL
- https://www.latent.space/feed
Summary
- Score
- 4.0
- Created
- 10 Oct 2026, 8:40 AM
- Tags
- Audience
- ai_ml_learnersdevelopers
What happened
In a Latent Space podcast panel published Oct 10, 2026, Google DeepMind's Pushmeet Kohli and Biohub's Sal Candido (moderated by Brandon Anderson) argue AlphaFold's protein structure prediction was only the beginning and did not solve protein folding, citing protein dynamics, disorder, and the limits of static structure prediction. They also claim scaling compute and data alone won't solve biology, that finding the right scaling law matters more than blindly increasing model size, and that low-quality metagenomic data can improve protein language models, which may already contain scientific knowledge researchers haven't unlocked.
Why it matters
For builders evaluating AI-for-science tooling or claims, this is a direct caution: static structure prediction like AlphaFold is not the same as modeling protein dynamics or whole cells, so don't treat it as a solved pipeline. The concrete claim worth testing is that low-quality metagenomic data can improve protein language models and that better data beats more data — relevant if you are choosing between scaling a model and curating a dataset. Beyond bio-AI niches, this is a conceptual discussion with no product, price, or API change to act on.
Discussion angle
If scaling compute and data isn't enough for biology, what does that imply for teams in other domains deciding between a bigger model and better data curation — and where does that tradeoff actually show up in your own project?