PairSAE cracks open protein AI's black box with interpretable features

Foundation models for structural biology, such as AlphaFold and Boltz-2, achieve remarkable accuracy in predicting biomolecular structures, but understanding why they make specific predictions remains a challenge. Standard sparse autoencoders (SAEs) work well on transformer-style sequence embeddings, but fail on pairformer-like architectures used in protein co-folding. Naively applying SAEs to pairwise representations causes a quadratic explosion of features and misses concepts that are distributed jointly across sequence and pair representations.

To solve this, researchers from the institutions of Migliorini et al. introduce PairSAE, which first summarizes pairwise tensors using an N-mode SVD to extract token-wise interaction roles. A sparse autoencoder then learns a shared set of token-level features that decode into both sequence and pair representations. Evaluated on Boltz-2 activations for PLINDER protein-ligand complexes, PairSAE produces interpretable features that align with UniProt functional annotations and can even predict Boltz-2's affinity values. This work, accepted at the MLSB 2025 workshop, provides a practical path to open the black box of structural biology AI, with direct implications for rational drug design and protein engineering.