Identifying drugs that reverse disease-associated transcriptomic features has been a well-established strategy in drug repurposing, yet its potential for de novo drug discovery remains largely untapped. Xing et al. recently presented the GPS (Gene Expression Profile Predictor on Chemical Structures) platform in Cell, which employs deep learning to directly predict compound-induced transcriptomic changes from chemical structures and enables virtual profiling of nearly seven million compounds for reversal score-driven screening. In hepatocellular carcinoma and idiopathic pulmonary fibrosis, GPS successfully identified novel compound series with favorable cellular selectivity and in vivo efficacy, demonstrating the feasibility of a closed-loop pipeline from virtual screening to lead optimization. This commentary systematically dissects GPS""s core innovations—the RCL(Robust Collaborative Learning) framework and Z-RGES (Z-score normalised Reverse Gene Expression Score), and benchmarks them against contemporaneous models including TranSiGen and MiTCP, delineating their complementary strengths and shared bottlenecks. Furthermore, we discuss three key frontiers—cross-context generalization assessment, multimodal validation integrating structural biology, and ADMET-incorporated multi-objective optimization—that shall guide the paradigm shift from repurposing known drugs to generative de novo design in transcriptomics-driven drug discovery.