Microbiome-based disease prediction with multimodal variational information bottlenecks.

Area of Science:

• Microbiome research
• Computational biology
• Machine learning in health

Background:

• The gut microbiome plays a crucial role in human health.
• Machine learning models can distinguish healthy from diseased microbiome states.
• Current methods analyze microbial species or strain data separately.

Purpose of the Study:

• To develop a novel deep learning model, the Multimodal Variational Information Bottleneck (MVIB).
• To integrate multiple gut microbiome data types (species and strain markers) for improved disease prediction.
• To create a faster and more interpretable predictive framework.

Main Methods:

• Developed the Multimodal Variational Information Bottleneck (MVIB), a deep learning model.
• Utilized an information-theoretic approach for joint representation learning.
• Applied the model to analyze shotgun metagenomic data from 11 disease cohorts.

Main Results:

• Achieved high disease prediction performance (0.80 < ROC AUC < 0.95) on 5 out of 6 disease cohorts.
• Demonstrated competitive classification performance and faster training times (< 1.4 seconds).
• Successfully identified key microbial species and strain markers contributing to predictions.

Conclusions:

• MVIB offers a powerful and efficient approach for multimodal microbiome data analysis.
• The model provides interpretable insights into disease-associated microbial features.
• MVIB demonstrates scalability and generalizability across different studies and data modalities.