InterPLM

Protein language models predict structure and function well enough to have reshaped computational biology, but what they actually represent internally has stayed opaque. Individual neurons are a poor place to look: models pack far more concepts than they have dimensions, so each neuron fires across many unrelated contexts. InterPLM, from Elana Simon and James Zou at Stanford University, attacks this by training sparse autoencoders (SAEs) on the residual stream of ESM-2 and reading the resulting sparse features rather than the raw neurons.

The result is a large, concrete vocabulary. Across layers, the SAEs yield up to 2,548 features that align strongly with known biological concepts, covering up to 143 distinct Swiss-Prot concepts — binding sites, structural motifs, functional domains, biophysical properties. Probing ESM-2's own neurons the same way finds at most 46 aligned neurons per layer spanning just 15 concepts, direct evidence that the model stores most of what it knows in superposition.

InterPLM is one of three independent, concurrent efforts to apply sparse dictionary learning to protein language models. It is distinct from InterProt (Adams and colleagues at Columbia and Ginkgo Bioworks), which uses TopK sparsity and targets structure and localization determinants, and from PLM-SAE (Shanghai Smart Logic Technology), which turns SAE features into zero-shot variant effect prediction. The three converge on a similar idea from different directions and are worth reading together rather than as successive versions.

#Key Features

• Feature dictionaries per layer: Autoencoders are trained separately on each of six layers of both backbones, so a user can compare what a concept looks like early versus late in the network rather than treating the model as one representation.
• Quantitative concept evaluation: Feature activations are binarized at several thresholds and scored against 433 amino-acid-level Swiss-Prot concepts with a domain-aware precision/recall, giving a reproducible metric for how interpretable an SAE actually is.
• Automated feature description: Because curated concepts label under 20% of features, a Claude-driven pipeline writes descriptions from maximally activating examples, then validates them by using the description to predict activations on held-out proteins.
• Missing-annotation discovery: Apparent false positives frequently turn out to be real. A Nudix-box feature and a peptidase S1 domain feature each flagged proteins with no Swiss-Prot label whose annotation InterPro independently confirms.
• Causal steering: Clamping a feature's activation on one residue shifts the model's amino-acid predictions at neighbouring positions, confirming that features drive behaviour rather than merely correlating with it.

#Technical Details

SAEs were trained on hidden representations from 5 million random UniRef50 sequences, with <cls> and <eos> tokens excluded and tokens shuffled within 1,000-protein shards. For the 8M backbone the expansion factor is 32x, mapping 320-dimensional embeddings into a dictionary of 10,240 features; the 650M autoencoders keep the same 10,240 dictionary over 1,280-dimensional embeddings. Each layer's SAE was selected from 20 candidates trained for 500,000 steps at batch size 2,048, sweeping learning rates from 1e-4 to 1e-8 and L1 penalties from 0.07 to 0.2. Features are normalized to a 0–1 range using maximum activations over 50,000 Swiss-Prot proteins so that scales are comparable across features. Scaling matters: the 650M model yields substantially more concept-aligned features than the 8M model, and clustering features by their dictionary vectors separates, for instance, three TonB-dependent-receptor detectors whose F1 scores range from 0.998 down to 0.611 purely by specificity.

#Applications

The immediate audience is anyone trying to understand or debug a protein model: feature dictionaries make it possible to ask which learned concepts drive a prediction, to spot spurious correlations, and to compare backbones on interpretability rather than benchmark accuracy alone. For curators, the annotation-gap results suggest a practical workflow — scan a proteome with a concept-aligned feature and triage the unlabeled hits. The InterPLM.ai dashboard exposes every feature of the 8M backbone for browsing without any local setup, and the released autoencoders plug into an existing ESM-2 pipeline in a few lines.

#Impact

Published in Nature Methods after circulating as a preprint, InterPLM helped establish sparse dictionary learning as a standard tool in protein model interpretability and supplied the concept-F1 evaluation that later work reuses. The code and pretrained autoencoders are released under the MIT license, and the training walkthrough is written to be adapted to other backbones. Limitations are inherent to the approach: concept evaluation is bounded by what Swiss-Prot curates, LLM-generated descriptions correlate only weakly with concept-classification performance (Pearson r = 0.11) and so measure a different axis of interpretability, and steering was demonstrated on short motif contexts rather than as a general sequence-design mechanism.