Skip to content

Protein-LM explainability — a review

Adjacent methods resource for the FM-to-Virtual-Cells talk. Hunklinger & Ferruz, "Towards the explainability of protein language models", Nat Mach Intell 2026.

What it is

A review of interpretability and explainability methods for protein language models — attention analysis, probing classifiers, sparse autoencoders (SAEs), and the rest of the toolkit for asking what a protein FM has actually learned. It is the protein-side survey of the same question the talk raises for single-cell FMs.

How it connects to the talk

The talk's mechanistic-interpretability thread — the teal cluster in the paper map — is the satisfying ending of Act 1: the mechanism behind the reckoning. The key sc-FM finding (Simon & Zou) is that single-cell FMs encode cell-type and pathway features but minimal regulatory logic.

That whole wave started on the protein side — Adams et al. 2025 PNAS applied SAEs to protein FMs first, and the sc-FM SAE papers followed. This review is the protein-side state of the art for the same toolkit:

  • It is the methods reference for the interpretability cluster — what techniques exist, what they can and can't show.
  • It is the comparison baseline — protein FMs are the most-studied biology FMs for interpretability, so "what works on protein LMs" sets expectations for what should work on sc-FMs.
  • It reinforces the talk's point that interpretability is now a family-spanning research program, not a single-cell curiosity.

Why it's adjacent, not core

A review, and protein-FM-focused rather than single-cell — so it is a resource page, not a paper-map node. But it is the right citation for anyone running the Track 1 mechanistic-interpretability project: it tells you which interpretability methods have a track record before you point them at a cancer sc-FM.

Read more

Last updated 2026-05-14.