Figure 2B source data files

APR-results.csv - Area under precision-recall (APR) curve for each dimensionality reduction method across microbiome datasets with varying data densities. Results are reported over ten cross-validation folds, each generated using a different train-test split.

AUC-ROC-results.csv - Area under the receiver operating characteristic curve (AUC-ROC) classification performance for each dimensionality reduction method across microbiome datasets with varying data densities. Results are reported over ten cross-validation folds, each generated using a different train-test split.

Mahalanobis-dist.csv - Mahalanobis distances between samples and their corresponding group centroids (IBD vs. non-IBD) for each dimensionality reduction method across microbiome datasets with varying data densities. Results are reported over ten cross-validation folds, each generated using a different train-test split.

PERMANOVA-pseudo-F.csv - PERMANOVA pseudo-F statistics (IBD vs. non-IBD) for each dimensionality reduction method across microbiome datasets with varying data densities. Results are reported over ten cross-validation folds, each generated using a different train-test split.

Notes: 
1. Benchmarked dimensionality reduction methods include Joint-RPCA, MOFA+, iClusterPlus, intNMF, and multiblock sPLS from mixOmics.
2. Microbiome data densities are encoded as follows:
	meta_g_taxonomic_profiles = original metagenomics dataset (12% density)
	meta_g_taxonomic_profiles_11 = 11% density
	meta_g_taxonomic_profiles_9 = 9% density
	meta_g_taxonomic_profiles_7 = 7% density
	meta_g_taxonomic_profiles_5 = 5% density
	meta_g_taxonomic_profiles_3 = 3% density
3. MOFA+ contains 4x more entries than the other methods because projection of the test set into the learned latent space was performed separately for each data modality (MetaG, Virome, MetaT, and Metabolomics). Labels corresponding to each modality-specific projection are provided in the source data for Figure S2.