MVIB

MVIB learns joint stochastic encodings from species-relative abundances and strain-level markers derived from shotgun metagenomic sequencing to predict human disease states.

Key Features:

• Multimodal Integration: MVIB combines species-relative abundances and strain-level markers derived from shotgun metagenomic sequencing into a single predictive framework.
• Deep Learning Architecture: It employs an information-theoretic deep neural network to compute a joint stochastic encoding of heterogeneous input modalities.
• Performance and Efficiency: Reported Receiver Operating Characteristic Area Under Curve (ROC AUC) values range from 0.80 to 0.95 across disease cohorts, with average training times under 1.4 seconds.
• Interpretability: The model uses a saliency technique to identify the microbial species and strain-level markers most relevant to predictions.
• Scalability and Generalization: MVIB can incorporate additional modalities such as metabolomic data from mass spectrometry and shows cross-study generalization comparable to Random Forest when trained and tested on different cohorts of the same disease.

Scientific Applications:

• Disease prediction and biomarker discovery: Analysis of human gut metagenomic samples from 11 publicly available disease cohorts spanning six different conditions to predict diseases and identify associated microbial markers.

Methodology:

Training deep neural networks using a variational information bottleneck to maximize task-relevant information while minimizing irrelevant details and computing a joint stochastic encoding, with outputs interpreted via a saliency technique.