Adversarial validation of causal identifiability assumptions in model-based inference of diseases from multi-level omics and clinical data

Model-based causal inference techniques are being used to infer disease mechanisms from heterogeneous biomedical data. However, the identifiability assumptions made in such models remain empirically unvalidated. We first formulated disease mechanism models as causal Bayesian networks (BNs) learned by integrating genomics, epigenomics (DNA methylation), transcriptomics (RNA-seq), proteomics Sequential Window Acquisition of All Theoretical Fragment ion Mass Spectra (SWATH-MS), metabolomics (gas chromatography-mass spectrometry (GC-MS)), and clinical phenotype data using constraint-based and score-based structure learning. Next, we developed an adversarial generator network trained to introduce latent confounders or violate causal sufficiency. The original and perturbed datasets were passed through the disease models to quantify changes in the inferred causal structure. We applied our approach to breast cancer models inferred from The Cancer Genome Atlas (TCGA) database. The adversarial network effectively generated perturbed data capturing latent confounding between genetic mutations and protein expression and feedback loops violating causal sufficiency. Retraining causal models on perturbed data altered causal structures, endorsing the limitations of the original identifiability assumptions. Our framework, therefore, appears to be a suitable tool for empirically evaluating the robustness of disease mechanism models when the core causal assumptions are violated. The adversarial validation framework is effective in evaluating the robustness of causal assumptions for identifying model-based disease mechanisms inference.