This research introduces DrugReflector, a deep learning architecture designed to accelerate phenotypic drug discovery by linking disease biology with chemical interventions through transcriptomics. By utilizing a lab-in-the-loop active reinforcement learning framework, the system iteratively refines gene signatures to prioritize compounds likely to induce desired cellular changes. In experimental trials involving hematopoietic stem cells, this computational approach achieved hit rates over ten-fold higher than traditional random screening methods. The study further demonstrates the framework's versatility by successfully identifying potential treatments for oncology indications and uncovering new biological pathways, such as the role of cholesterol biosynthesis in megakaryocyte development. Ultimately, this methodology bridges the gap between high-resolution single-cell data and scalable drug testing to identify complex molecular modulators.

References:

• DeMeo B, Nesbitt C, Miller S A, et al. Active learning framework leveraging transcriptomics identifies modulators of disease phenotypes[J]. Science, 2025, 390(6776): eadi8577.