An Integrative Multi-Omics Random Forest Framework for Robust Biomarker Discovery

Abstract High-throughput technologies now produce a wide array of omics data, from genomic and transcriptomic profiles to epigenomic and proteomic measurements. Integrating multiple omics layers measured on the same samples can reveal cross-layer molecular hubs that single-layer analyses miss. We present an unsupervised, multivariate random forest (MRF) framework with an inverse minimal depth (IMD) importance to prioritize shared biomarkers across omics. In each forest, one layer serves as a multivariate response and the other as predictors; IMD summarizes how early a predictor (or response MSRV) appears across trees, yielding interpretable, cross-layer feature rankings. We provide three IMD-based selection strategies and introduce an optional IMD power transform to enhance sensitivity to interaction signals. In extensive simulations spanning linear, nonlinear, and interaction regimes, our method matches SPLS/CCA under linear settings and outperforms them as nonlinearity increases, while adapted univariate ensemble learners (RF, GBM, XGBoost) underperform in the multivariate, unsupervised context. Applied to TCGA BRCA and COAD, MRF-IMD identifies genes, CpGs, and miRNAs enriched for cancer-relevant pathways and yields more robust survival stratification than linear integrators with matched model sizes. In a TCGA pan-cancer analysis, MRF-IMD features achieve higher ARI than alternatives and recover coherent tumor-type clusters; in ADNI, the integrative signature improves dementia-progression stratification over a published methylation risk score. Our scalable, interpretable MRF-IMD framework advances reliable multi-omics biomarker discovery when nonlinear, cross-layer dependencies matter.