The enzymatic activity of the microbiome toward carbohydrates in the human digestive system is of enormous health significance. Predicting how carbohydrates through food intake may affect the distribution and balance of gut microbiota remains a major challenge. Understanding the enzyme/substrate specificity relationship of the carbohydrate-active enzyme (CAZyme) encoded by the vast gut microbiome will be an important step to address this question. In this study, we seek to establish an in silico approach to studying the enzyme/substrate binding interaction. We focused on the key CAZyme and established a novel Poisson noise-based few-shot learning neural network (pFSLNN) for predicting the binding affinity of indigestible carbohydrates. This approach achieved higher accuracy than other classic FSLNNs, and we have also formulated new algorithms for feature generation using only a few amino acid (AA) sequences. Sliding bin regression is integrated with minimum redundancy maximum relevance for feature selection. The resulting pFSLNN is an efficient model to predict the binding affinity between CAZyme and common oligosaccharides. This model can be potentially applied to the binding affinity prediction of other protein/ligand interactions based on limited AA sequences.
Keywords: CAZyme; Poisson process; few-shot learning; fructo-oligosaccharides; machine learning; neural network; resistant sugar; sugar binding affinity.