Machine learning-based prediction of cross-immunity
Abstract Cross-immunity, defined as the ability of T-cells to recognize multiple antigen peptide-major histocompatibility complexes, is a fundamental feature of adaptive immunity. However, the prediction of different peptide epitopes that can be recognized by the same T-cell receptor remains challenging. Currently, artificial intelligent (AI)-based machine learning (ML) methods can be successfully used for pattern recognition in epitope molecular space by detecting the functional similarity between peptide sequences. In this study, using literature-based experimental data, we examined ML-based binary classification models trained on small datasets to predict the activity of nine-amino-acid-long peptides. Our results suggest that the consensus function of well-established similarity matrix-based representations and structural-based descriptors of epitopes yields better performance because representation-specific noises are reduced and individual model weaknesses are partially compensated. We also sought to determine the extent to which the predictive power of the applied AIs procedure depended on the physicochemical content of the descriptor set during the training process. In addition, challenging the models, we applied them to an independent experimental dataset to examine the effects of diverse laboratory conditions on a regulated biological measurement. In summary, applying a consensus function can capture the biological complexity of cross-reactivity at the binary classification level, even when applied to relatively small datasets.