A deep learning framework for disease classification integrating microbiome and metabolome: explaining multi-omics contribution

Multi-omics data analysis combines different and complementary information content to provide novel insights into several bioinformatics scenarios. In this context, an important objective of microbiomics and metabolomics is to study the complex interaction between microbiome and metabolome and their relationships with host disease. Deep learning ofers a suitable framework for enabling the straightforward and simultaneous processing of complex, multidimensional omics data. This paper presents a computational framework comprising a deep learning model and an explainability module. The deep model, called microBiome-metaBolome Network (BiBoNet), performs disease classification through an integrated analysis of both microbiomics and metabolomics data. Employing an explainability technique, namely SHAP, we try to shed light on the network’s behaviour and quantify the relative contribution of each omics data type to the model’s output. We validate BiBoNet against other classical machine learning algorithms, considering five different matched microbiome/metabolome datasets related to different diseases. In all cases, BiBoNet reached the best results in terms of several score metrics.