Graph convolutional network with self-attention pooling for the prediction of neutralizing paratope sequences of SARS-CoV2 antibodies

Abstract

The COVID-19 pandemic caused by the Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV2) pathogen has resulted in a great loss to human lives and economic disrupton. Although the severity of the disease outbreak has been overcome and normal operatons have resumed in many countries, therapeutcs to treat COVID-19 stll remain necessary as many in the populaton contnue to get re-infected with circulatng variants of the SARS- CoV2 pathogen. It would be ideal to have a repertoire of suitable antbody or paratope sequences which can be rapidly designed for therapeutc needs, based on emergent strains.

In-silico models provided by deep graph networks are an avenue for high-throughput discoveries of neutralizing antbody sequences. Graph neural networks have emerged as promising architectures in several aspects of health and molecular medicine, such as in adaptve graph relatons for antbody predicton, [1] models of drug-target interactons [2] and to aggregate spatally related cellular data [3]. Here, a deep graph neural network employing graph convoluton with self-atenton pooling was trained to detect pairs of neutralizing paratopes and epitopes from sequence data alone.