DeepHeal: Consensus Variational Autoencoder Atlases of Drug-Induced Proteomic Perturbations Enable Phenotype-Based Drug Discovery and Mechanism-of-Action Analysis

Drug-perturbation proteomic profiles encode mechanism-relevant phenotypes, but robust latentspace models for large-scale perturbation proteomics remain underdeveloped. Here we developed DeepHeal, a framework comprising complementary consensus variational autoencoder atlases derived from 875 compound-induced proteomic log2FC profiles in HCT116, and used it for phenotype-based drug discovery, mechanism-of-action analysis and candidate prioritization. To our knowledge, this is the first application of variational autoencoders to large-scale drug-perturbation proteomics. Our primary atlas combined supervised contrastive learning with a 600-model consensus VAE ensemble so that the full cohort contributed to representation learning while 448 labelled compounds imposed class-aware structure; a matched unsupervised average-profile atlas served as reference. The supervised atlas retained 257 consensus latents and achieved perfect within-cohort separation of antineoplastic and nonantineoplastic compounds with both LGMLVQ and the best-performing AutoGluon classifier identified for the supervised branch (KNeighbors), whereas the average-profile atlas retained 194 latents and was weaker for classification (the best-performing AutoGluon classifier identified for the average-profile branch (TabM) accuracy 0.767, AUROC 0.756; LGMLVQ accuracy 0.767, AUROC 0.738). Pathway interpretation revealed a trade-off between density and breadth: the supervised atlas concentrated signal in cell-cycle, replicationstress and DNA-damage programmes and recovered 100 canonical pathways overall, whereas the averageprofile atlas preserved broader normalized pathway diversity per 100 latents in both canonical and oncogenic collections. Applying a stringent four-model consensus rule to 427 previously unlabelled compounds identified 19 high-confidence candidates, including 14 rediscovered external antineoplastics and a smaller set of mechanistically plausible repurposing candidates led by JP1302, BAY 61-3606 and SB 225002. Cluster-level and macro-mechanism analyses further resolved cohort-wide and subgroup-specific pathway structure across the atlas. These results establish DeepHeal as a practical framework for perturbation proteomics and show that proteomic latent atlases can support phenotypic screening, mechanism-of-action analysis and hypothesis generation within large drug-response resources.