Filtering viral data from protein models cuts viral performance, but fine-tuning restores it for under $80
See how this story connectsThe UK AI Security Institute’s Chem-Bio team studied pretraining data filtering on the open-weight ESM-2 protein language models, pretraining models from 8M to 3B parameters with viral protein sequences removed. Filtering lowered performance on viral proteins while largely preserving performance on non-viral proteins, with effects across masked language modelling, mutation-effect prediction and contact prediction that varied by task and model size. Fine-tuning on the withheld viral data fully restored the lost performance, at about $80 of H100 compute for the largest 3B-parameter model.
Why it matters
AISI says data filtering does not protect open-weight protein models on its own, because the lost capability can be recovered cheaply, and that it needs to be paired with measures such as data access controls.
- When
- 26 Oct 2026
- ID
WTA-N-2026-0435- Where
- United Kingdom
- Who
- UK AI Security Institute
Summary written with AI by wheresthe.ai, from UK AI Security Institute. Checked against the source on 26 Oct 2026.
Source copy: download the page as saved on 26 Oct 2026 (PDF). Also held by the Internet Archive.
Cite this story
WTA-N-2026-0435: “Filtering viral data from protein models cuts viral performance, but fine-tuning restores it for under $80.” wheresthe.ai, 26 Oct 2026. https://wheresthe.ai/d/gb-filtering-viral-data-from-protein-models-cuts-viral-perfo-2026-10/ (source: UK AI Security Institute).
Spotted an error? Email [email protected].