A Michigan State University team is betting that artificial intelligence can speed up the hunt for new medicines, and early lab results suggest they might be onto something. Using a custom AI system, researchers identified chemical compounds that show early promise against two conditions that lack effective treatments: hepatocellular carcinoma, the most aggressive form of liver cancer, and idiopathic pulmonary fibrosis. In preclinical tests, lead molecules reduced tumor size in mice and lowered fibrosis markers in human lung tissue. Those findings appeared this month in the journal Cell, and the group also released the model's code plus a public web portal so other labs can run their own virtual screens.
Michigan State's College of Human Medicine laid out the work in a March 17 press release, describing how the platform, dubbed GPS for "Gene expression profile Predictor on chemical Structures," was trained on millions of experimental measurements to predict whether a compound will increase or decrease the activity of specific genes. The release quotes lead researcher Bin Chen and several collaborators on the cross-disciplinary push and highlights clinical partners who supplied tissue samples for validation tests.
How the GPS platform works
The peer-reviewed Cell paper, listed on PubMed, describes GPS as a deep learning model that predicts transcriptomic perturbation signatures directly from chemical structures, then uses optimization methods to rank and refine candidate compounds. The authors report that GPS learns structure-gene-activity relationships and performs large virtual screens to flag molecules that appear likely to reverse disease-associated gene expression patterns.
Validated hits: liver cancer and lung fibrosis
In hepatocellular carcinoma models, the team reports two novel compound series with favorable cellular selectivity that reduced tumor size in mice. For idiopathic pulmonary fibrosis, the work surfaced one repurposing candidate and a new anti-fibrotic molecule that lowered fibrosis markers in human explant lung tissue obtained through a clinical collaboration. The Cell article walks through the in vitro and in vivo validation steps and emphasizes that these hits still need medicinal-chemistry optimization and safety testing before anyone talks about human trials, per ScienceDirect.
Open tools and next steps
The team has made GPS's source code public and launched an online portal where other researchers can run virtual screens and review model outputs for themselves. The paper and project materials point to the GitHub repository and to OCTAD for interactive screening and follow-up.
Local viewers got a layperson-friendly primer when FOX 2 Detroit aired a video interview with Dr. Bin Chen on April 2, where he walked through the GPS approach and the early lab results. That coverage builds on broader reporting and university materials, but industry statistics offer a reality check. A recent analysis of clinical development found that from 2011 to 2020, the overall likelihood that a program entering Phase I would reach approval was under 10 percent, underscoring how rarely preclinical promise turns into an approved drug.
