AI-Guided Structure-Aware Modeling and Thermal Proteomics Reveal Direct Demethylzeylasteral-ACLY Interaction

Abstract

Identifying the direct molecular targets of bioactive natural products remains a central challenge in chemical biology. Here we present an integrated experimental-computational framework, that combines matrix-augmented thermal proteomics with HoloGNN, a holistic graph neural network, to systematically prioritize and validate protein-ligand interactions. Benchmarking with PDBbind datasets HoloGNN achieves state-of-the-art performance. Applying this framework to 50 structurally diverse natural products identified Demethylzeylasteral as a direct interactor of ACLY. Orthogonal biochemical assays confirmed micromolar binding and enzymatic inhibition. In an imiquimod-induced psoriasis-like mouse model, Demethylzeylasteral reduced disease severity and inflammatory cytokine expression. Single-cell transcriptomics revealed that Demethylzeylasteral reverses keratinocyte hyperproliferation and suppresses ACLY-dependent lipid metabolic reprogramming. Together, this scalable, closed-loop strategy integrates thermal proteomics and machine learning to uncover direct targets of natural products and provides mechanistic evidence linking ACLY inhibition to therapeutic modulation of inflammatory pathology.