Human-in-the-Loop Targeted Molecule Design Informed by Transcriptomes

Hang Yuan, Chen Li, Wenjun Ma, Yuncheng Jiang

Keywords: Targeted molecule design; Human-in-the-loop; Transcriptome-guided; Molecular generation

Abstract: Transcriptome-guided targeted molecule design contributes to the advancement of precision medicine. However, the fragmented and poorly integrated nature of existing datasets limits the ability of transcriptome-aware approaches to meet the design expectations of chemists, often requiring additional expert intervention after molecular generation. To address this gap, we propose HiTGen, a human-in-the-loop targeted molecule generation framework informed by transcriptomes. Specifically, HiTGen operates in two phases: i) Transcriptomes are served as the central biological driver and are fused with expert a posteriori knowledge via a tailored bidirectional attention mechanism, enabling biochemically grounded guidance for diffusion-based generation of molecule candidates. ii) An expert-guided human-in-the-loop optimization mechanism is employed by HiTGen to refine these candidates toward desired molecular targets. Extensive experiments demonstrate that HiTGen consistently outperforms state-of-the-art models across ten evaluation metrics, yielding chemist-aligned targeted molecules with potential for precision medicine. The code will be released upon acceptance of the paper.

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

Submission Number: 7538