Abstract
Therapeutic monoclonal antibodies (mAbs) have garnered significant attention for their efficacy in treating a variety of diseases. However, some candidate antibodies exhibit non-specific binding to off-target proteins or other biomolecules, leading to high polyreactivity, which can compromise therapeutic efficacy and cause other complications, thereby reducing the approval rate of antibody drug candidates. Therefore, predicting the polyreactivity risk of therapeutic mAbs at an early stage of development is crucial. In this study, we fine-tuned six pre-trained protein language models (PLMs) to predict the polyreactivity of antibody sequences. The most effective model, named PolyXpert, demonstrated a sensitivity (SN) of 90.10%, specificity (SP) of 90.08%, accuracy (ACC) of 90.10%, F1-score of 0.9301, Matthews correlation coefficient (MCC) of 0.7654, and an area under curve (AUC) of 0.9672 on the external independent test dataset. These results suggest its potential as a valuable in-silico tool for assessing antibody polyreactivity and for selecting superior therapeutic mAb candidates for clinical development. Furthermore, we demonstrated that fine-tuned language model classifiers exhibit enhanced prediction robustness compared with classifiers trained on pre-trained model embeddings. PolyXpert can be easily available at https://github.com/zzyywww/PolyXpert.