{Xuetao Cao Chunmei Wang Yu Chen Meihan Zhang Zhu Liang Xiaohang Leo Fang Jian-Qing Zheng Hantao Lou}

Abstract

The COVID-19 pandemic has been ongoing for nearly two and half years, and new variants of concern (VOCs) of SARS-CoV-2 continue to emerge, which urges the development of broadly neutralizing antibodies. Variants such as the delta (B.1.617.2 lineage) and Omicron (BA.1 and BA.2) were reported to exhibit immune evasion to some of the current therapeutic antibodies. The ever-evolving SARS-CoV-2 calls for rapid prediction of antibody binding to new variants and development of broadly neutralizing antibodies. Considering the application of deep learning in antibody engineering and optimization, we wonder whether the broadly reactive antibodies against SARS-CoV-2 variants can be rapidly designed and generated by deep learning. Here we report the development of an Atrous Convolution Neural Network (ACNN) based deep learning framework: cross-reactive B cell receptor network (XBCR-net) that can predict broadly reactive antibodies against SARS-CoV-2 and VOCs directly from single-cell BCR sequences. XBCR-net composes of two parts, the first part extracts the features relevant to the antibody–antigen interaction via three-branch ACNN, and the second part predicts the binding probability of the antibodies to antigens (14 different RBD sequences) by a residual structural Multi-Layer Perceptron. The performance of the ACNN-based XBCR-net prediction on SARS-CoV-2 binding was evaluated, showing significantly higher accuracy, precision and recall value than other frameworks.