AlphaFold Didn't Solve Protein Folding

Summary

In the Latent Space panel, Pushmeet Kohli from DeepMind and Sal Candido from Biohub argue that AlphaFold's static structure prediction is merely a starting point. They emphasize that achieving true biological understanding requires solving protein dynamics, disorder, and whole-system modeling. They contend that simply increasing computational power and data volume will be insufficient without the right inductive biases and data quality.

Why it matters

The discussion challenges the common narrative that a single groundbreaking model has 'solved' a field. It reminds the AI and biotech communities that protein folding is just one part of a much harder problem: predicting how living systems behave over time. For those investing in AI-driven drug discovery or synthetic biology, the emphasis on finding the correct scaling laws and the right data—rather than just piling on FLOPs—has direct implications for where research funding and engineering efforts should be directed.

Highlights

• AlphaFold predicts a single static structure. Protein dynamics, conformational disorder, and multi-protein interactions remain largely unsolved.
• The existence of scaling laws in biology is not guaranteed. Identifying the correct data distribution and architecture is as important as raw computational power.
• Low-quality metagenomic sequences can paradoxically improve protein language models. However, chasing easily accessible data risks optimizing for practicality rather than scientific impact.
• Biohub's framework: treating diseases requires 10x–100x leaps, not 10% incremental gains. This shifts the expectation threshold for what 'AI for biology' should deliver.

Critical take

The panel is inherently promotional, with both guests marketing their respective institutions' missions. Consequently, the insistence that AlphaFold has not 'solved' protein folding risks downplaying a tool that has transformed structural biology and accelerated thousands of experiments. Additionally, the transcript cuts off after about three minutes of a 32-minute discussion. As a result, the most comprehensive technical claims (cryo-EM representations, virtual cell datasets, uncertainty calibration) are merely listed as topics and never explored in depth. The reader is left with a strong thesis but limited evidence.