Microbiome-based disease prediction with multimodal variational information bottlenecks

Abstract

Scientific research is shedding more and more light on the interaction of the gut microbiome with the human body and its role in human health. Empirical evidence shows that the microbiome is strongly interconnected with the host immune system and can contribute to carcinogenesis. Furthermore, studies show that it can be used as a predictor for various diseases. In this work, we propose the Multimodal Variational Information Bottleneck (MVIB), a novel machine learning model capable of learning a joint representation of multiple heterogneous data modalities. Adopting an information theoretic interpretation of deep neural networks, MVIB allows to compute a joint stochastic encoding of the input data modalities, which acts as a minimal sufficient statistics of the inputs for the prediction of a target label. In a supervised learning setting, we use the MVIB to predict whether patients are affected by a certain disease by jointly analising the species-relative abundance and strain-level marker profiles extracted from shotgun metagenomic sequencing of microbial DNA. We propose various pre-processing methods for the marker profiles and validate our model on ten disease datasets. Additionaly, we perform transfer learning experiments by fine-tuning the model on the target disease dataset after a pre-training on all other source diseases. Furthermore, we adopt a saliency technique to interpret the outcome of the model classification and identify which microbial species and strain-level markers mostly contributed to a certain prediction. We show that MVIB achieves competitive results on the microbiome-based disease prediction task. The code for this work is publicly available at https://github.com/nec-research/microbiome-mvib.