Integrated multi-omics prioritizes small-molecule candidates for the N-cadherin HAV site and reveals endpoint-level toxicity trade-offs

Abstract N-cadherin (CDH2) is a central effector of epithelial-to-mesenchymal transition (EMT) that contributes to invasion, metastasis and therapy resistance. Although peptide and antibody approaches validate N-cadherin as a therapeutic target, small-molecule modulators of its adhesive interface remain limited. We aimed to prioritize candidate compounds that suppress EMT-associated transcriptional programs, show plausible binding at the N-cadherin HAV (His79-Ala80-Val81) hotspot, preserve the biomass objective in context-specific metabolic models, and minimize predicted endpoint-level toxicity liabilities. LINCS L1000 level-5 profiles were screened for concordant down-regulation of CDH2, SNAI1 and SNAI2, yielding 20 compounds across cell lines, doses and exposure times. Compounds were docked to the N-cadherin EC1 HAV pocket and compared with a spatially separated non-HAV EC1 control surface. Toxicity was evaluated using a multi-task DeepChem Graph Convolutional Network trained on Tox21 with a scaffold-based split. Predicted toxicity risk was summarized using max-risk, defined as the highest predicted positive-class probability across 12 endpoints. Context-specific metabolic effects were modeled with Human1 using E-Flux, pFBA and FVA, and summarized using Objective Preservation Ratio (OPR), Metabolic Impact Score (MIS), Targeted Pathway Selectivity (TPS), and Homeostasis Disruption Score (HDS). Evidence pillars were winsorized, normalized and integrated using a transparent weighted composite score. Docking energies ranged from − 8.1 to − 5.9 kcal mol –1 , and 19 of 20 compounds showed more favorable predicted binding at the HAV pocket than at the non-HAV control surface. The toxicity model achieved a micro-averaged ROC-AUC of 0.775 (bootstrap 95% CI 0.756 – 0.794) and PR-AUC of 0.335 (bootstrap 95% CI 0.298 – 0.374). Across modeled conditions, median OPR was 1.12, median TPS was 0.165, median MIS was 0.79, and HDS was uniformly low. Under the conservative max-risk-based integration, BRD.K09587429 ranked first (composite score 0.656). It showed concordant suppression of CDH2, SNAI1 and SNAI2, a HAV-site docking score of − 6.7 kcal mol –1 , OPR of 1.081, TPS of 0.184, and max-risk of 0.554 at SR-ARE. BRD.K04021425 ranked second and was retained as the alternative candidate, with the strongest docking score and the lowest endpoint-level max-risk. A 100-ns molecular dynamics simulation showed persistent computational occupancy of the HAV-site region, with His79 and Ala80 remaining within 0.4 nm of the ligand in 96.9% and 99.1% of frames, respectively, over the 20–100 ns interval. MM-PBSA analysis of 81 snapshots yielded a favorable end-state binding-energy estimate of − 24.58 ± 6.90 kcal mol – 1 , which remained consistent when analysis was restricted to the final 50 ns ( – 27.22 ± 5.68 kcal mol –1 ). This multi-omics framework prioritizes candidate N-cadherin/EMT modulators by jointly considering transcriptomic directionality, HAV-site docking, endpoint-resolved toxicity, and metabolic selectivity. BRD.K09587429 emerged as the highest-ranked discovery-stage candidate, with molecular dynamics showing persistent computational occupancy of the HAV-site region and MM-PBSA indicating a favorable model-dependent end-state energetic profile; however, direct binding, functional N-cadherin antagonism, molecular mechanism, and the predicted SR-ARE-associated liability require experimental validation. The framework distinguishes structural plausibility from safety and can be adapted to other protein-protein interfaces relevant to metastatic cancer biology.