#17 · ML / AI methods

Interpretable models expose folding, concept, and rewiring structure in biology

The dominant thread is making learned models of biological systems structurally interpretable. A sparse autoencoder analysis of ESMFold's folding trunk separates the roles of single and pair representations and traces concept features to fold quality. CUTTIE clusters neural network units into a tree and localizes predictive concepts through top-down conditional independence tests with family-wise error control. Co$^2$ODE learns sparse neural or symbolic ODEs with a conditional mask and VAE closure to capture context-dependent rewiring in biochemical networks. What is new is attention to mechanism-level structure, whether internal features of a protein language model, units of a general neural network, or the parameters of a biochemical ODE. The three items address distinct problems and methods, so the shared element is interpretability tied to biological mechanism rather than a common technique.