Abstract
The identification of linear B-cell epitopes (LBCEs) plays a pivotal role in the development of immunological products such as immunodiagnostic, vaccines and therapeutic antibodies. This criticality has led to the development of computational approaches for the prediction of LBCEs, aiming at accelerating discovery and prioritising targets for experimental assessment. Existing LBCE predictors rely on having access to large volumes of data from a wide range of pathogens to train generalist predictive models, which can result in biases towards widely studied organisms and compromise performance when focusing on neglected or emerging pathogens. We introduce EpitopeTransfer, a phylogeny-aware transfer learning strategy that results in noticeable gains in expected predictive performance. This is achieved by fine tuning large protein language models using abundant data from higher-level taxa and applying this refined feature embedder for fitting pathogen- or lower taxon-specific models. We report substantially increased performance in comparison to state-of-the-art approaches for LBCE prediction across a broad spectrum of pathogens (viruses, bacteria and eukaryotes), and show that these gains can be directly attributed to the use of phylogeny-aware fine tuning of the feature embedder coupled with pathogen- or taxon-optimized modelling.