NbBayesLM: bayesian prediction of nanobody thermostability using protein language model

Nanobodies, single-domain antibodies derived from camelids, are promising biologics due to their small size and high stability. Accurate prediction of their thermostability is critical for therapeutic and diagnostic applications. Due to their ability to bind conformationally constrained epitopes that are typically inaccessible to conventional antibodies, nanobodies represent a uniquely valuable modality in therapeutic target engagement and drug discovery. Existing methods for predicting nanobody thermostability often rely on limited data, handcrafted features, or black-box machine learning models that lack uncertainty quantification, limiting their generalizability and reliability. To address these gaps, this study, named NbBayesLM, proposes a Bayesian neural network (BNN) approach that integrates protein language model (PLM) embeddings with chemical property features to predict nanobody thermostability. In our formulation, physicochemical properties are incorporated as Bayesian priors, providing biologically meaningful constraints that guide posterior learning and improve model interpretability. Trained on a dataset of 10,630 nanobody sequences with experimentally determined $T_m$ values, our model achieves a mean absolute error of 1.89 °C and $R^2$ score of 0.67, outperforming existing models reported in the literature, while the fusion mechanism enhances performance over unimodal approaches and the BNN architecture provides well-calibrated uncertainty estimates to guide candidate selection and accelerate nanobody engineering.