Interpretability

What models internally encode and how they use it: attention analysis, layer-wise probing, sparse autoencoders, and mechanistic explanations. Generalization tests remain under evidence/generalization.

Quartz 4

Explorer

Interpretability

ESMFold has massive activations that disproportionately contribute to output quality

PLMs learn family-specific protein contacts from sequence context windows of about 20-40 amino acids

Protein language models learn structure-level features, including disorder, in later layers

Sequence-only protein language models implicitly cluster proteins at fineness levels that increase with size

Quanta hypothesis

AF3-generation methods incorporate MSA information exclusively into the pair representation

Protein language models are in-context learners that default to retrieval over reasoning

Sparse autoencoders recover protein-family and gene-ontology features from PLM representations

Linear probes trained on SAE features from intermediate steps of protein structure prediction neural networks can be used for property prediction

Attention matrices from antibody LMs can be used for paratope prediction

Attention matrices of antibody language models do not correspond to contacts

Categorical Jacobian method

Evo-velocity

Light chain representations affected by heavy chain, but not vice versa

Membrane proteins are predicted by PLMs via solvent-exposed hydrophobic residues

Ordered SAEs are more steerable than top-k SAEs

PLM attention maps from specific heads can be used to predict allosteric networks

PLM attention matrices correspond to 3D contacts

PLMs cluster antibodies from the same repertoire by V gene and maturation status

Protein folding neural networks do not learn the physics of protein folding for most proteins

Protein folding neural networks make local optimizations in the absence of coevolutionary information

Protein language models can predict zero-shot which proteins belong to the same species

Sequence-based DL models learn intra-family phylogenetics of protein evolution

Sparse autoencoder-derived features do not outperform PLM-derived embeddings for downstream prediction

Sparse autoencoder

Subnetworks within protein language models encode specific protein families

Likelihood of predicted contacts by MSA transformer weakly correlate with dwell time in molecular dynamics simulations

The Evoformer can calculate ligand-binding residues

Transformers attend to gap tokens to "opt out" of updates