AlphaFold was a major advance in predicting protein structures, but a structure prediction is not a general solution to protein folding—or a predictive model of biology. In this conversation, Pushmeet Kohli and Sal Candido unpack that distinction and the technical work still required to model living systems.
They examine why more compute and more data are not sufficient by themselves: biological datasets are uneven, experimental measurements are costly and context-dependent, and it is not yet clear which scaling laws apply. That makes biology different from domains where large, relatively uniform datasets can support straightforward scaling. The discussion also probes a consequential design choice: when should researchers encode scientific intuition into a model, and when should they favor general-purpose architectures that might discover patterns researchers did not anticipate?
For practitioners, the value is in how the guests frame the research problem rather than in a promise of imminent breakthroughs. Useful predictive models must connect structure to function and account for the conditions under which biological systems operate. That requires better data and experiments, careful evaluation, and a clearer understanding of what a model is being asked to predict. The episode offers a grounded way to assess claims about AI for biology: ask what the model predicts, what evidence supports that prediction, and whether success on a defined benchmark transfers to the complexity of living systems.
Key Insights
- AlphaFold’s achievement in predicting protein structures is not equivalent to solving the full protein-folding problem or predicting how proteins behave in every biological context.
- Scaling compute and dataset size alone is not a reliable recipe for progress when biological data is heterogeneous, costly to generate, and dependent on experimental conditions.
- Finding scaling laws in biology is difficult because datasets and measurements vary across systems; researchers cannot assume the regularities seen in more uniform AI domains will transfer.
- Model designers face a real tradeoff between incorporating scientific intuition and using general-purpose architectures that may learn patterns researchers did not specify.
- Predicting living systems requires connecting molecular structure to function and context, not merely producing a plausible structural output.
- Evaluating AI-for-biology progress requires asking whether performance on a bounded benchmark transfers to the biological predictions researchers and experimentalists actually need.
Who should listen: AI-for-biology researchers, biotech R&D leaders, and technical investors assessing whether foundation-model scaling can produce experimentally useful biological predictions.
Why This Matters
The frontier question is shifting from whether AI can produce impressive biological representations to whether it can make reliable, testable predictions about living systems. For builders and investors, the episode is a useful corrective to treating compute and model scale as substitutes for experimental design, data quality, and domain-specific evaluation.
