From Genomic Whispers to Therapeutics: Multi-Resolution Transcriptome-Guided Diffusion Models for Drug Design and Screening

ZIYU XU, zijian zhang, Liang Wang, Zhiyuan Liu, Qiang Liu, Shu Wu, Zhaoxiang Zhang, Liang Wang

Keywords: Transcriptome-Guided Drug Design, Single-Cell Transcriptomics, Molecular Generation, Graph Diffusion Models

Abstract: Traditional drug discovery is protracted and extremely expensive. While Structure-based Drug Design (SBDD) has advanced AI-driven molecular generation, target-centric models struggle with diseases arising from the dysregulation of complex physiological systems. To bridge this gap, we introduce Transcriptome-based Drug Design (TBDD): designing molecules from a cell’s transcriptomic response to perturbations. We present scTrans-Gen, a diffusion model that conditions generation on multi-resolution transcriptomic data (bulk and single-cell). Central to our approach is a transcriptome-centric condition extractor that aligns perturbation signals across domains into a function-oriented chemical space, avoiding the ill-posed reconstruction of microscopic structures from macroscopic signals. To exploit single-cell data, we propose a Gene Pseudoimage mechanism for robust high-resolution conditioning. Across diverse benchmarks, scTrans-Gen outperforms strong baselines on multiple metrics. We further demonstrate novel inhibitor design for specified gene knockouts and an efficient generate-then-search screening workflow suitable for time-sensitive clinical scenarios. Altogether, scTrans-Gen offers a practical route to function-oriented drug discovery and personalized precision medicine.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 3130