Paired-unit activation steering for context-aware protein design

PULSE improves context-aware activation steering for protein language models, achieving higher optimization success rates and final predicted property scores across solubility, immunogenicity, and toxicity objectives.

• Why it matters: Global steering vectors ignore that property-improving directions differ by protein sequence background, limiting controllable design when context shifts across protein populations.
• What they did: PULSE matches positive- and negative-property proteins by maximizing representation cosine similarity, then builds steering vectors from similarity-weighted, unit-normalized differences across matched pairs. It is compared against global population-centroid steering, and LDS is introduced to assess alignment of iterative representation changes with property-improving directions from similar backgrounds.
• The result: Across three objectives, PULSE consistently outperforms global centroid steering in both predicted property scores and optimization success rates, while LDS enables predictor-free candidate prioritization by capturing trajectory reliability differences.

AI summary of the abstract · may contain errors · not medical advice · read the paper.