Uncovering Features of ProteinMPNN with Sparse Autoencoders

Sparse autoencoders integrated with ProteinMPNN uncover correlations between sparse latent activations and protein features, including secondary structure, residue volume/shape, disulfide bridges, and ligand coordination.

• Why it matters: ProteinMPNN’s latent representations are hard to interpret, and polysemanticity is hypothesized to underlie this; mapping activations to concrete protein properties can enable mechanistic understanding and control.
• What they did: The authors train sparse autoencoders to expand ProteinMPNN’s latent space and analyze how activations in sparse dimensions correlate with multiple protein feature types.
• The result: By manipulating sparse activations, they generate a Brolucizumab short chain variable fragment variant that is more thermostable than variants produced by traditional high temperature molecular dynamics–based methods.

AI summary of the abstract · may contain errors · not medical advice · read the paper.