Knowledge-Guided, Causally Constrained Multi-Omic Graph Learning for Patient-Specific Signaling-Driver Inference in Breast Cancer

Background: The transcriptomic subtypes of breast cancers can have distinct signaling states because RNA levels do not adequately reflect the levels of proteins and phosphoproteins. Although cohort-level prediction-focused classifiers are prevalent in the field of graph-based multi-omic classifiers, directionally constrained and patient-level network scores are necessary for clinically interpretable prioritization. Methods: Sample-level harmonization of matched RNA sequencing, reverse-phase protein array (RPPA), and clinical matrices was performed. The main analysis comprised of 742 primary tumours, 20,530 transcript measurements, and 131 RPPA analytes. This resulted in 103 nodes representing genes being mapped with RPPA analytes. A structural-equation model was constrained by a 132-edge acyclic signaling prior spanning model that included programs for 13 different pathways: receptor tyrosine kinase, PI3K-AKT-mTOR, MAPK, hormone-receptor, DNA-damage, apoptosis, Hippo, WNT, TGF-beta, adhesion, and cell-cycle. After the heteroskedasticity-robust testing and Benjamini-Hochberg correction, prior-sign agreement was performed, and 200 bootstrap resamples were used to keep the edges. The patient specific putative-driver scores were composed of role-oriented protein activity, structural residual activation, and total downstream influence. The classification of PAM50 was evaluated on 673 tumours; cross-validation was done five times on each tumour, with leakage control. The associations between overall-survival and variables were tested in 582 tumours and 70 deaths. Results: RNA abundance correlated significantly with 111 of 131 RPPA analytes at a false-discovery rate <0.05. The median RNA-protein correlation for total proteins, cleaved proteins, and phosphoproteins was 0.370, 0.375, and 0.143, respectively, with both total and cleaved proteins showing significantly greater levels of correlation than phosphoproteins (Spearman rho 0.370; 0.375; 0.143; P=3.04x10^-5 and 1.12x10^-4). Of the 132 previous edges, 52 were both directionally concordant and FDR-significant and were also bootstrap-stable. Supported relationships included ESR1-GATA3 (standardized  = 0.481), AKT1-AKT1S1 (  = 0.533), MAP2K1-MAPK1 (  = 0.497), MET-RAF1 (  = 0.444), CHEK2-TP53 (  = 0.448), and MTOR-RPS6KB1 (  = 0.297). The early-fusion RNA-RPPA model achieved balanced accuracy 0.867 (95% CI 0.863-0.872) and macro-AUROC 0.975 (95% CI 0.973-0.976). Recall was 0.958 for Basal, 0.836 for HER2-enriched, 0.906 for Luminal A, and 0.765 for Luminal B disease. ERBB2 regained control at the patient level on HER2-enriched tumours, INPP4B lost control on Basal tumours and ESR1-associated programs were regained on luminal disease. There was no pathway level score that survived multiplicity correction that had an association with overall survival. Conclusions: Stable and directionally constrained signaling graph and heterogeneous patient-specific driver profiles were supported by matched transcriptomic and proteomic data. The findings illustrate the interpretive benefits of proteomic integration, as well as highlighting that cross-sectional TCGA data enable causally constrained inferences, but do not allow for definitive causal identification. External validation and genomic perturbation anchors remain necessary before clinical deployment.