Phenotype prediction using biologically interpretable neural networks on multi-cohort multi-omics data

This study developed interpretable predictive models for multi-omics data using biologically informed neural networks (visible networks). The models' performance, interpretability, and generalizability were evaluated on four population cohorts (Ntotal = 2940) from the BIOS consortium, predicting smoking status, age, and LDL levels using RNA expression and CpG methylation data. High performance was consistently achieved for smoking status prediction (mean AUC = 0.95), with interpretability revealing known genes like AHRR, GPR15, and LRRN3. LDL level prediction generalized well only in one cohort, and age prediction showed a mean error of 5.16 years. Multi-omics networks consistently outperformed single-omics models in terms of performance, stability, and generalizability.

Arno van Hilten, Jeroen van Rooij, M. Arfan Ikram, Wiro J. Niessen, Joyce B. J. van Meurs, Gennady V. Roshchupkin