In a Latent Space panel, Google DeepMind's Pushmeet Kohli and Biohub's Sal Candido argued that AlphaFold's breakthrough was only the beginning. Both agreed that scaling compute and data alone is not enough to solve biology. The real work, Candido said, is finding the situations where more compute and data actually produce better results—and that depends on having data with the right information and statistics.
Kohli noted that AlphaFold's handcrafted architecture reflected scientific intuition, but protein folding is far from fully solved. Static structure prediction leaves out protein dynamics and disorder, and the panel pointed to cryo-EM micrographs as a potentially richer representation. Candido added that biological modeling must move beyond individual proteins to whole systems, and that building a virtual cell will require fundamentally different datasets.
The two speakers largely agreed throughout the discussion, with no substantive disagreement appearing in the source. They also suggested that protein language models may contain hidden scientific knowledge, and that trustworthiness and uncertainty calibration matter more than full interpretability. The panel framed the path to curing disease as needing 10x breakthroughs rather than incremental improvements.