RoDPAA: A Robust Dissimilarity-Based Principal Angle Analysis Framework for Multi-View Microbiome Data Integration

RoDPAA, an innovative framework for integrating multi-view microbiome data, has been developed to address challenges posed by sparse, compositional data, limited sample sizes, and methods that only recognize exact sharing. The researchers introduce a dissimilarity-based approach, utilizing data-appropriate dissimilarities such as Bray-Curtis without the need for log-ratio transformation or likelihood assumptions.

This unique method involves aligning view-specific signal subspaces through principal angles, capturing both complete and partial sharing, and testing shared components via permutation analysis. The beauty of RoDPAA lies in its simplicity, requiring only eigendecomposition and minimal tuning. In a series of Gaussian simulations, RoDPAA demonstrated its prowess by matching the performance of established frameworks JIVE and MOFA while running faster and recovering shared dimensions consistently across replicates.

When applied to two inflammatory bowel disease (IBD) cohorts characterized by sparse, zero-inflated microbiome data, RoDPAA's Bray-Curtis-based approach successfully revealed shared structures that remained stable even after prevalence filtering. In stark contrast, log-ratio-based methods failed to yield similar results. Perhaps most intriguingly, the leading microbiome-metabolome factors identified by RoDPAA, learned without any disease labels, emerged as the sole factors associated with IBD.