Semipres: an Integrated Deep Learning Method of Sequence and 3d Object Detection Models for Host-microbe Interactions by Identification of Protein-protein Contact Residues

Background: Interface mimicry is an important strategy for microbes to efficiently interfere with host protein-protein interactions to maximize their survival and spread. With interface mimicry, microbial proteins with surface residues that can structurally mimic the known interactor of host proteins have a potential to interact with the host. Computationally predicting such interactions is critical to understand the mechanisms of microbe-induced diseases and drug discovery. Computational methods that do not use protein 3D structures (e.g., sequence-only methods) are generally less accurate, and those that require 3D structures are limited by the structural coverage and cannot utilize the abundant information available from sequences. Results: Here we present a new deep learning method that integrates ALBERT, a natural language processing model with PointNet, a 3D object identification model. The ALBERT component learns from the vast amount of available sequence data, and the PointNet component makes it available to predict the contact residues. We show that a natural language processing model can combine with PointNet and be applied to the interface mimicry concept for predicting protein-protein interactions. On a benchmark, the method outperformed a graph convolutional neural network model with a similar purpose. Conclusions: Our deep learning method for protein-protein interaction prediction is the first method that integrates sequence-based and structure-based deep learning to predict interface mimicry. Our analyses indicate that it performs better for heterodimer predictions than homodimers. This integration is a promising idea, however the method can further be optimized for better predictive performance.