ToxinMI

Accurately identifying peptide toxicity is a crucial step for computer-aided peptide-based drug screening, which could accelerate novel drug discovery and reduce resource consumption. Recently, deep learning has shown promising performance in bioinformatics. However, one challenge in developing a deep learning-based model for peptide toxicity prediction is how to represent peptides effectively. In this study, we propose an end-to-end deep learning model named ToxinMI, to predict peptide toxicity that learns features directly from sequence alone. Precisely, ToxinMI captures the sequential and evolutionary features of the peptide simultaneously and introduces the mutual information principle to learn a discriminative representation by discarding noisy information and retaining related-task information from them as much as possible. The experimental results demonstrate that ToxinMI achieves superior predictive performance against state-of-the-art baselines.1