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

In a round‑table hosted by Latent Space, Pushmeet Kohli from Google DeepMind and Sal Candido of the Biohub discussed the limits of AlphaFold and the future of AI in biology. Both agreed that AlphaFold’s success is only the beginning of a larger process: simply scaling compute and data is not enough to grasp the complexity of biological systems. They emphasized that identifying scaling laws specific to each data type is more important than merely enlarging models, and that the quality of information—such as metagenomic data or cryo‑electron micrographs—is crucial for training protein language models that can capture dynamics, disorder, and other non‑static properties.

Kohli recalled that AlphaFold’s architecture blended scientific intuition with custom design, suggesting that inductive biases remain necessary. Candido argued that current models must move from predicting isolated structures to representing complete virtual cells, which will require fundamentally different datasets and a “10× break‑through” mindset rather than incremental gains. They also debated the importance of confidence and uncertainty calibration over full interpretability, and the possibility that future AI systems may understand other AIs better than humans do.

This conversation is significant because it outlines the next challenges for AI in medicine and drug discovery, where major advances could accelerate treatment identification and ultimately help achieve the goal of curing diseases.

Read the original article on Latent Space

This summary is an informational synthesis produced by dataqbs.com. All rights to the original content belong to its author and the cited media outlet. We act solely as curators of technology news and claim no authorship.