#13 · ML / AI methods

Sparse autoencoders steer pretrained protein and genomic language model representations

The common thread is adapting or interrogating pretrained biological models rather than building task-specific networks from nothing. COMPASS-PTM builds on a protein language model to jointly predict residue-level PTM sites and enzyme-substrate assignments. PLM-SAE steers sparse autoencoder features of ESM embeddings to improve zero-shot variant effect prediction across 114 deep mutational scanning datasets, and Hidden State Genomics applies sparse autoencoders and feature knowledge graphs to Nucleotide Transformer v2, validated with interventions. What is new is mechanistic access to internal representations. Other items, including FlexENN, GISTnet-MD and DMG-PPI, apply graph neural networks to binding, dynamics and interaction networks, and the reported support is primarily benchmark-based.