AlphaFold's Limits: Why AI Still Hasn't Solved Protein Biology

In a panel discussion, Pushmeet Kohli from Google DeepMind and Sal Candido from Biohub challenge the assumption that AlphaFold solved protein folding. They argue that simply scaling compute and data falls short—the real breakthrough requires finding the right scaling laws, high-quality datasets, and integrating scientific expertise with AI architecture.

The conversation centers on lessons from building AI for biology: why protein structure prediction remains unsolved, how protein language models may already contain undiscovered scientific knowledge, and why modeling living systems requires rethinking how AI researchers approach data and inductive biases. Candido emphasizes that curing disease demands 10x improvements, not incremental gains, while Kohli reflects on balancing human intuition with general-purpose AI models.

They explore the gap between optimizing for available data and optimizing for the most important scientific questions—a tension that will shape the next generation of AI-driven biology.