Transformer-based Deep Learning Empowers the Discovery of Self-Assembling Peptides with over Ten Trillion Sequence Quantities

Abstract Self-assembling of peptides is essential for a variety of biological and medical applications. However, it is challenging to investigate the self-assembling properties of peptides within the complete sequence space due to the enormous sequence quantities. Here, we demonstrate that a Transformer-based deep learning model is effective in predicting the aggregation propensity (AP) of peptide systems, even for decapeptide and mixed pentapeptide systems with over ten trillion sequence quantities. Based on the predicted AP values, we are not only able to derive the aggregation laws for designing self-assembling peptides, but also reveal the transferability relation among the AP's of pentapeptides, decapeptides, and mixed pentapeptides, leading to discoveries of self-assembling peptides by concatenating or mixing, as consolidated by experiments. Our deep learning approach enables speedy, accurate, and thorough search and design of self-assembling peptides with the complete sequence space of oligopeptides, advancing peptide science by inspiring new biological and medical applications.