Paired-unit activation steering for context-aware protein design

Abstract

Activation steering in protein language models enables controllable sequence generation without changing the model weights. However, existing methods typically construct steering vectors from global differences between property-positive and property-negative protein populations, overlooking how property-improving directions vary across different protein sequence backgrounds. To address this, we propose \textit{Paired-Unit local steering of embeddings (PULSE)}, which matches positive- and negative-property proteins by maximizing their representation cosine similarity and constructs steering vectors by aggregating similarity-weighted, unit-normalized differences between matched pairs. Across three protein optimization objectives--solubility, immunogenicity, and toxicity--PULSE consistently outperforms global population-centroid steering in both final predicted property scores and optimization success rates. Further analysis reveals that conventional static steering-vector diagnostics cannot adequately explain optimization performance. To characterize steering reliability during iterative sequence editing, we introduce \textit{Local Directional Support (LDS)}, which measures how well successive representation changes align with property-improving directions from similar protein backgrounds. We show that LDS captures differences in trajectory reliability and enables predictor-free candidate prioritization, enriching successful protein designs without requiring property predictions during candidate selection. Together, these findings demonstrate the importance of context-aware steering and trajectory-level evaluation for controllable protein sequence optimization.