InterProt

Protein language models predict structure and function well, but nobody can say precisely what they have learned. Individual neurons in a model like ESM-2 are polysemantic — one unit fires on unrelated concepts — so inspecting them directly yields little. Sparse autoencoders address this by decomposing model activations into a much wider, sparsely activating dictionary in which individual latents tend to correspond to single interpretable concepts, a technique that has become standard for probing large language models.

InterProt, from the Department of Systems Biology at Columbia University with a co-author at Ginkgo Bioworks, applies that technique to ESM-2 and adds the missing piece: a tool for actually reading the results. The released package is a family of TopK sparse autoencoders trained on the residual stream of ESM-2 650M, plus an open web visualizer that shows, for any latent, the sequences that activate it most and where on their structures the activation falls. The team finds that many latents are crisply interpretable — secondary structure elements, conserved motifs, disulfide contexts, entire domains — and that a large subset fire almost exclusively on single protein families, evidence that ESM-2 carries an internal notion of sequence homology.

The work was published as a spotlight poster at ICML 2025. It is distinct from the concurrent InterPLM, which trained ReLU-based SAEs on ESM-2 and evaluated them automatically against Swiss-Prot annotations; InterProt uses TopK SAEs, a blinded human interpretability study, and linear probing on downstream tasks. A comparable approach on other backbones appears in PLM-SAE.

#Key Features

• TopK sparsity: Only the k largest latents are allowed to be non-zero, which sets the L0 norm of the encoding directly and gives better reconstruction at a given sparsity than penalty-based alternatives.
• Latent taxonomy: A rule-based scheme classifies each latent by activation pattern — point, periodic, motif, or domain — and by family specificity, defined as reaching an F1 above 0.7 when used alone to classify InterPro family membership.
• Interpretable probing: Linear probes trained on SAE latents match ESM probes on downstream tasks while making their own coefficients readable, so the top-weighted latents can be inspected to explain a prediction.
• Open visualizer: interprot.com aligns the most-activating sequences for a latent and paints activation strength onto the corresponding structure, the interface used to run the paper's human rating study.

#Technical Details

The SAEs use a TopK activation on a linear encoder and decoder trained with plain reconstruction MSE, and were fit on activations from one million random UniRef50 sequences under 1,022 residues. The headline checkpoint is a layer-24 SAE with hidden dimension 4,096 over ESM-2 650M's 1,280-dimensional residual stream, with SAEs for the other layers also released. Sweeping hyperparameters shows that lowering k or raising the expansion factor both increase the number of family-specific latents, while activation-pattern classifications stay stable; family specificity peaks in early-to-middle layers and falls off later. In a blinded study, seven raters familiar with protein biology scored 100 randomly drawn latents each, and roughly 80% of SAE latents were rated interpretable against a far lower rate for ESM neurons. Probes were evaluated on secondary structure, subcellular localization, thermostability, and CHO-cell expression; SAE probes track ESM probes across layers and beat them on secondary structure. Inspecting the top coefficients recovers real biology — bipartite nuclear localization signals matching R/K(X)10-12KRXK, signal-peptide cleavage sites, and a hydrophobic-core latent that is the most positive predictor of thermostability.

#Applications

InterProt is for anyone who needs to know why a protein language model made a prediction: protein engineers auditing a fitness predictor, model developers diagnosing why early layers underperform on a task, and biologists mining pLM internals for hypotheses about sequence determinants that have no annotation yet. The linear-probing workflow is the practical core — train a probe on SAE latents rather than raw embeddings, then read its largest weights as a shortlist of candidate mechanisms to test.

#Impact

InterProt helped establish sparse autoencoders as a working tool in protein machine learning rather than a language-model-only technique, and its family-specificity finding gives concrete support to the view that pLMs lean on homology rather than biophysics. The released artifacts are unusually complete for interpretability work — code under the MIT license, weights on Hugging Face under Apache 2.0, and a live visualizer — which has made the checkpoints a common starting point for downstream SAE analyses. Honest limits remain: interpretations rest on human judgment of top-activating examples, many predictive latents still have no known functional association, and steering ESM through family-specific latents is demonstrated but not yet useful.