How IBDex was built
A skimmable overview. For full methodological detail, see the manuscript and source repository.
1. Data
3,168 adult intestinal bulk RNA-seq samples harmonised across four IBD cohorts (GSE193677, IBDome, HMP2, GSE165512). Gene identifiers were mapped to HGNC symbols, and expression was normalised as log2(CPM + 1). Samples were split by patient into train/validation/test (70/15/15), stratified by disease, so no patient appears in more than one split.
2. Gene panel selection
A defensible 3,000-gene panel was selected by a multi-evidence weighted score across STRING, Reactome, KEGG, GO-BP, Open Targets, GWAS and a landmark gene set, keeping disease-relevant genes while removing batch- and tissue-driven noise.
3. The conditional variational autoencoder
A CVAE with gradient-reversal adversaries is trained to suppress dataset
and fine tissue-site identity in its 16-dimensional latent space, while
retaining disease and inflammation signal. The frozen encoder is what
the explorer and the ibdex-projector CLI both use to place
a sample into that space.
4. Endotype discovery
Four colon (C1–C4) and four ileum (I1–I4) endotypes were discovered by consensus clustering over four base algorithms, run separately per tissue given their distinct baseline biology.
5. Pathway scoring
Decoder-native, tissue-specific pathway scores are computed directly from the latent space and calibrated against the reference cohort's distribution, validated against ssGSEA but distinct from it.
6. External projection
New samples, including entirely external cohorts, are projected into the frozen geometry without retraining. All dataset condition indicators are set to zero, matching the setting where the incoming cohort is not one of the four training datasets.
ibdex-projector
command-line tool are available at
github.com/ajithak9945/IBDEX_Package.