8/ Paper: pnas.org/doi/10.1073/...
MIT News article: news.mit.edu/2025/researc...
Github:

7/ We show that SAE & transcoder features are much more interpretable than ESM neurons, for both protein-level & amino acid-level representations. This has the potential to improve safety, trust & explainability of PLMs. As PLMs improve, SAEs could help us learn new biology.

6/ We also use Claude to autointerpret SAE features based on protein names, families, gene names & GO terms. Many features correspond to families (like NAD Kinase, IUNH, PTH) & functions (like methyltransferase activity, olfactory/gustatory perception).

5/ We interpret these SAE features using Gene Ontology (GO) enrichment. Many protein-level SAE features align tightly with GO terms across all levels of the GO hierarchy.

4/ SAEs have a very wide latent dimension with a sparsity constraint. This forces PLM representations to disentangle into biologically interpretable, sparsely activating features without any supervision.

3/ We train sparse autoencoders (SAEs) on protein-level and amino acid-level representations from layers 6-10 of ESM2_t12_35M_UR50D. We also train transcoders (an SAE variant) on protein-level representations.

2/ Protein-level representations from PLMs are used in many downstream tasks. Disentangling their features can enhance interpretability, helping us trust and explain downstream applications.

1/ PLMs like ESM have made big strides in predicting protein structure & function. But they feel like a “black-box.” What biological information do PLM representations contain? Can we disentangle them systematically?

Excited to share our recent work: Sparse autoencoders uncover biologically interpretable features in protein language model representations now in PNAS. Thread below 🧵

Hello world! See the thread below for our recent work 🌿 MINT!