Stability Priors Enable Few-Shot Variant Effect Prediction under Limited Label Budgets

Protein variant prioritization is a recurring challenge in drug discovery, from optimizing therapeutic proteins to identifying target mutations that alter activity or binding, yet campaign-specific experiments often provide only tens to hundreds of labeled variants. We present PsiFit, a stability-informed framework for low-$N$ fitness prediction that integrates mutation-induced stability changes predicted by a multimodal sequence--structure foundation model into contrastive fine-tuning of protein language models (pLMs). Across 114 non-stability ProteinGym assays with $N{=}96$ labeled variants, PsiFit outperforms both zero-shot ESM-1v and a strong supervised low-$N$ baseline, showing that structural stability provides a transferable biophysical prior even when the measured phenotype is not stability itself. PsiFit provides a step toward data-efficient, experimentally actionable variant prioritization under limited assay budgets.