Why AlphaFold Didn't Solve the Protein Folding Problem
In a Latent Space podcast interview, Google DeepMind's Pushmeet Kohli and CZ Biohub's Sal Candido discussed why AlphaFold has not completely solved the protein folding problem. They explored how AI scaling principles must be adapted to capture complex biological dynamics beyond static structure prediction. Differentiating static 3D structure prediction from dynamic protein folding kinetics is critical for advancing cell biology and targeted drug discovery. It highlights that applying general AI scaling laws to biology requires moving beyond existing static databases toward novel data generation. While AlphaFold predicts the static equilibrium 3D structure of proteins from amino acid sequences, it does not simulate the kinetic pathways, intermediate states, or dynamic environmental interactions of folding in real time. The experts noted that leveraging Rich Sutton's 'Bitter Lesson' in biology requires generating richer dynamic datasets rather than simply increasing compute on existing structural data.
## BACKGROUND
The traditional protein folding problem asks how an unfolded amino acid chain physically folds into a functional three-dimensional shape over time. AlphaFold successfully solved structure prediction (identifying the final 3D shape), which is distinct from modeling the dynamic folding process itself. Additionally, Sutton's 'Bitter Lesson' in AI posits that general methods leveraging raw computation consistently outperform approaches engineered with hardcoded human domain knowledge.