AbTune: layer-wise selective fine-tuning of protein language models for antibodies.

Antibodies play central roles in immune defense and are widely used as therapeutic agents. However, the high structural and sequence diversity of antigen-binding loops, combined with limited experimental data and weak co-evolutionary signals, makes it difficult to develop generalizable predictive models. In this work, we investigate test-time fine-tuning strategies to improve protein language model (pLM) performance in low-data settings, with a focus on antibody-related tasks. Systematic evaluations across tasks show that carefully constrained fine-tuning greatly enhances performance while preserving generalization. In particular, depth-selective fine-tuning consistently outperforms full-depth fine-tuning, with optimal performance achieved when tuning 50%-75% of model layers for medium- to small-sized pLMs. We introduce AbTune, a test-time fine-tuning framework that leverages this depth-controlled adaptation strategy. Across antibody structure prediction, mutation effect prediction, and binding affinity prediction, AbTune outperforms both standard pLM baselines and task-specific predictors, achieving the best performance among the evaluated baselines on two of the three tasks. To gain insight into the adaptation process and identify optimal AbTune protocols, we analyzed representation shifts, examined how sequence properties influence fine-tuning dynamics, and evaluated metrics that capture potential overfitting. Our results show that fine-tuning depth, duration, and perplexity jointly influence performance and must be carefully controlled to achieve optimal results.