Striped-Cross Attention Network with Implicit Semantic Knowledge for Antibody Structure Prediction

Miao Gu,Min Liu

DOI: https://doi.org/10.1007/978-981-16-6775-6_29

2023-01-01

Abstract:The structure of an antibody directly determines its ability to bind with the target proteins specifically. It is of great significance to obtain accurate structure data for the antibody research in the field of disease diagnosis and therapy. Since antibody structure resolution experiments are time-consuming and costly, a growing number of methods for protein structure prediction are proposed to address this problem. Although considerable improvements have been made in the general protein structure prediction, deep learning-based approaches still fail to yield sufficiently accurate antibody structures to provide functional insights and design assistance. In this paper, we firstly exploit the unique semantic information implied from the comparison between the antibody structures and the general protein structures. We further propose a Striped-Cross Attention Network (SCA-Net) to efficiently fuse the features of different local functional regions and that of global regions on the antibody. Concretely, for each target amino acid position, SCA-Net collects its contextual information preferentially at the same functional domain (CDR or FR region) of the same peptide chain (heavy or light chain). Subsequently, the contextual information is fused with the global information. We utilize 2D ResNet with dilated convolution as the backbone network for feature extraction, and then construct a parallelly structured attention network using the proposed SCA-Net component as a classifier. Finally, an end-to-end multi-task learning framework is yielded to predict the antibody structures, which are described by the inter-residue distances and orientations. Quantitative experiments on two independent antibody structure test sets suggest that the proposed method achieves the expected results.