AET5: A transcriptome-guided molecular generation framework with contrastive self-supervised learning

Gene expression profiles capture system-level drug responses and offer a promising basis for de novo molecular generation. However, their application is limited by data sparsity and experimental noise, which hinder the reliable mapping between disease-associated transcriptomic perturbations and chemically valid therapeutic molecules. Here, we present AET5, a de novo molecular generation framework that conditions molecular design on disease-reversal gene expression profiles. AET5 integrates contrastive self-supervised learning with pre-trained sequence-to-sequence models to learn robust associations between transcriptomic signatures and molecular structures by deriving noise-tolerant transcriptomic representations and aligning them with molecular sequence space. Across the L1000 dataset, AET5 outperforms existing expression-guided generation methods in generation quality and distributional characteristics, while maintaining favorable physicochemical and drug-related properties. We further apply AET5 to generate candidate compounds for SARS-CoV-2 infection and prostate cancer. Molecular docking and dynamics simulations indicate stable target binding, supporting the biological relevance of the generated molecules. These results demonstrate that disease-reversal expression profiles can effectively guide de novo molecular generation, providing a general framework for biologically informed drug design under noisy transcriptomic conditions.Author summary: Gene expression profiles provide a system-level view of how cells respond to diseases and drug treatments, making them valuable for computational drug discovery. However, directly using transcriptomic data to guide de novo molecular generation remains challenging because these data are often sparse, noisy, and difficult to map reliably to valid chemical structures. In this study, we propose AET5, an expression-guided molecular generation framework that designs molecules conditioned on disease-reversal gene expression profiles. AET5 combines contrastive self-supervised learning with a pre-trained sequence-to-sequence molecular generator to learn robust associations between transcriptomic signatures and molecular structures, thereby reducing the influence of experimental noise and improving the alignment between biological perturbation information and chemical space. Experiments on the L1000 dataset show that AET5 achieves improved overall generation performance compared with existing expression-guided methods and produces molecules with favorable structural, distributional, and physicochemical characteristics. Case studies on SARS-CoV-2 infection and prostate cancer further suggest that the generated compounds may have biologically relevant target-binding potential. Overall, AET5 provides a scalable and biologically informed strategy for transcriptome-guided molecular design.