Work

DeepHeal: An Interpretable Framework for Phenotype-Based Antineoplastic Drug Discovery Using Drug-Perturbation Proteomics

Year: 2026

Type: preprint

Abstract: Drug-perturbation proteomics provides a direct molecular readout of compound-induced cellular states, but its high dimensionality and limited interpretability hinder its use for mechanism-oriented drug discovery. Here, we present DeepHeal, an interpretable consensus representation framework that transforms drug-induced proteomic responses into biologically readable and experimentally actionable antineoplastic response programs. Using an HCT116 perturbation atlas comprising 875 compounds and 9,960 protein features, DeepHeal integrates ensemble variational autoencoders, supervised contrastive learning, consensus latent-variable aggregation, and protein-level attribution. A label-independent Average branch captures recurrent proteomic variation, whereas a complementary label-aware SupCon branch concentrates known antineoplastic structure. The resulting latent programs are mapped back to proteins, pathways, and interaction networks, revealing shared cell-cycle, DNA-replication, stress-response, metabolic, proteostasis, apoptosis-related programs, together with mechanism-dependent response heterogeneity. DeepHeal prioritized 19 initially unlabeled compounds for experimental follow-up, of which seven inhibited HCT116 cells with IC₅₀ values below 10 μM, including two compounds with nanomolar activity. Without retraining, DeepHeal representations further recovered antineoplastic response states across 566 external proteomic perturbation profiles spanning multiple cellular contexts. Together, these findings establish DeepHeal as, to our knowledge, the first proteomics-first interpretable consensus representation framework that transforms large-scale drug-perturbation proteomes into protein- and pathway-resolved response programs for cross-context anticancer candidate prioritization and experimental validation.

Source: ChemRxiv

DOI: 10.26434/chemrxiv.15000952/v2