Week 14
Black box no more?
Also this week: Biomolecular systems modeling.
A step back from the deep learning weeks: can these models be trusted and understood? One thread builds physical constraints and domain knowledge into deep models, as AlphaFold does with equivariant geometry and AlphaFold 1 did with a learned potential plus optimization. The other asks models to explain their predictions, from attribution maps to probing what a protein language model has learned. A second part covers biomolecular systems modeling: network topology, steady states, dynamics, and the newer AI approaches to them. There are no assigned readings; the syllabus lists literature for both parts.
Objectives
- Name the places domain knowledge can enter a deep model (input features such as MSAs, architecture such as equivariant layers, loss terms such as steric-clash or bond-geometry penalties, and hybrid learned-potential-plus-optimizer pipelines) and give a protein example of each.
- Argue when built-in physical constraints help (little data, out-of-distribution inputs, physically valid outputs) and when they can hurt (wrong or oversimplified assumptions).
- Distinguish intrinsically interpretable models from post-hoc explanations of black-box models, and global explanations of the whole model from local explanations of one prediction.
- Compute and compare feature attributions for one prediction: gradient saliency, integrated gradients, perturbation methods such as occlusion, LIME, and SHAP, and attention weights, and state the pitfalls of each.
- Probe what a trained model encodes, for example by fitting linear probes to pLM embeddings or reading contacts from attention maps, and say what such a probe can and cannot show.
- Represent a biomolecular system as a network (genes, proteins, metabolites, reactions) and compute topological features: degree distribution, hubs, motifs, and modules.
- Model a small reaction or regulatory network with ODEs (for example mass-action kinetics), solve for its steady states, and check their stability from the eigenvalues of the Jacobian.
- Describe how newer AI methods enter systems modeling, such as learning dynamics from data with neural ODEs or physics-informed networks, and predicting cell responses to perturbations from single-cell data.