ESMFold’s pair representation carried more readable geometric information
Source check: Source checked. Not independent scientific confirmation.
Finding: A readout study found that ESMFold’s intermediate representations added up to 0.32 AUROC for recovering relationships within folded structures over the upstream ESM-2 representation. Single-position attributes were already largely recoverable upstream. The result distinguishes two kinds of information in the model.
Technology: ESMFold; ESM-2; fitted sparse autoencoders
Family ESM / ESMFold: profile · models page
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- Fitted sparse dictionaries exposed features from the trained folding network. Comparative readouts assessed where geometry-related information became more accessible.
- Test and result
- Learned dictionary readouts were compared with features from the upstream sequence model. The authors also checked whether changing interpretation prompts shifted the apparent division between representation types.
- Limitation / still unknown
- The finding concerns information recoverable from one trained folding model. It does not establish that the attributed directions are causal or describe how a protein folds in a cell.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.36226v1 · https://arxiv.org/html/2610.36226v1 · https://arxiv.org/pdf/2610.36226v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 21 Nov 2026. Review scope: §§3–4, readout and circuit-attribution checks; PDF front matter.
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 2/5 · Validation 2/5 · Novelty 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.