Accelerating Drug Discovery with “Paradigm Shifting” AI Model

Key Takeaways

• GPS predicts transcriptomic response from chemical structure alone
• Trained on millions of measurements across 70 cell lines
• Identified two novel HCC compounds and two IPF candidates
• Validated in human cells, animal models, and lung tissue
• Open-source code and web portal enable broader research use

Summary

A multi‑institution team led by Michigan State University unveiled GPS, a machine‑learning platform that predicts how a compound will alter gene expression from its chemical structure. Trained on millions of transcriptomic measurements across more than 70 cell lines, GPS screened ultra‑large libraries and pinpointed promising candidates for hepatocellular carcinoma (HCC) and idiopathic pulmonary fibrosis (IPF). The approach yielded two novel HCC compounds and both a repurposed and a novel anti‑fibrotic molecule for IPF, which were validated in human cells, animal models, and patient tissue. The researchers released the code and a web portal to enable broader adoption.

Pulse Analysis

Artificial intelligence is reshaping pharmaceutical research, yet most AI‑driven screens still rely on existing assay data, limiting exploration of truly novel chemistry. Transcriptomics offers a functional readout of disease pathways, but generating expression profiles for every candidate is prohibitively expensive. The GPS platform sidesteps this bottleneck by learning the relationship between molecular structure and gene‑expression signatures, allowing researchers to infer transcriptomic impact without wet‑lab experiments. This paradigm shift aligns computational efficiency with biological relevance, addressing a long‑standing gap in de novo drug discovery.

The GPS model was built on an unprecedented dataset: millions of public measurements covering 978 landmark genes across 70 cell lines, with a focus on four widely used lines (MCF7, HEPG2, PC3, VCAP). Leveraging this foundation, the team screened vast chemical libraries and identified two previously unknown compounds that reverse oncogenic signatures in hepatocellular carcinoma, as well as a repurposed drug and a novel anti‑fibrotic molecule for idiopathic pulmonary fibrosis. Each candidate demonstrated activity in relevant human cell lines, animal disease models, and, for IPF, authentic human lung tissue, providing a robust validation pipeline that bridges in silico prediction and preclinical efficacy.

Beyond the immediate therapeutic leads, GPS’s open‑source release and user‑friendly web portal democratize access to high‑throughput, transcriptomics‑guided screening. Pharmaceutical firms and academic labs can now integrate this tool into existing pipelines, accelerating target identification across diverse disease areas. As the industry seeks faster, cheaper routes to first‑in‑class medicines, platforms like GPS exemplify how AI can translate massive omics data into actionable drug candidates, potentially reshaping the economics and timelines of early‑stage discovery.