A team led by Michigan State University has used machine learning to identify promising treatments for two diseases with limited options: hepatocellular carcinoma, the deadliest form of liver cancer, and idiopathic pulmonary fibrosis, a chronic lung disease with a median survival of just three years.
The research, published March 17 in Cell, demonstrates how AI can accelerate early-stage drug discovery by predicting a compound’s biological effects from its chemical structure alone.
How GPS Works
The team developed a system called GPS—Gene expression profile Predictor on chemical Structures—that learns from millions of experimental measurements showing how different chemicals affect gene activity in cells.
“In our approach, instead of looking at cats or dogs, we want to know whether the compound will regulate up or down gene expression,” said Bin Chen, a senior author at MSU’s College of Human Medicine, drawing an analogy to how image recognition AI distinguishes animals.
Traditional drug discovery often starts with a known target—a specific protein to block or activate. GPS flips this approach. It predicts which compounds might reverse the abnormal gene expression patterns seen in diseased tissue, without needing to know the specific mechanism.
Jiayu Zhou, formerly at MSU and now at the University of Michigan, noted that biological data is inherently noisy. The model was designed to “learn from the data without being thrown off by all the noise.”
Results in Mice and Human Tissue
For hepatocellular carcinoma—the third leading cause of cancer death worldwide—the team screened compounds virtually using GPS, then tested the top candidates in mice with tumors. Two new compounds reduced tumor size.
For idiopathic pulmonary fibrosis, they identified one repurposed drug and two novel compounds that showed promise. These were validated not just in mice but also on human lung tissue samples from Corewell Health’s lung transplant program.
“Drug discovery is a team sport, and not for the faint of heart,” said Edmund Ellsworth, director of MSU’s Medicinal Chemistry Facility. The project involved more than 20 researchers across MSU, Stanford’s Asian Liver Center, the University of Michigan, and Corewell Health.
What This Means
The approach addresses a bottleneck in drug development. Testing every possible compound against every disease in the lab is impossible—there are too many of both. GPS lets researchers computationally narrow millions of candidates to a handful worth testing.
Samuel So and Mei-Sze Chua at Stanford’s Asian Liver Center noted the platform “greatly expands the pool of novel compounds” for diseases like liver cancer, where existing treatments often stop working.
The team has made their code and a web portal publicly available at apps.octad.org/GPS/, allowing other researchers to screen compounds for diseases of interest. This open approach means the method could be applied beyond the two diseases studied here.
The Fine Print
These are still early results. Compounds that work in mice often fail in humans—the historical success rate from Phase I trials to approval is under 10%. The two cancer compounds and three lung fibrosis candidates identified here need years of additional testing before they could reach patients.
The study also doesn’t address toxicity at therapeutic doses, drug-drug interactions, or how these compounds behave in humans. What GPS provides is a faster, more systematic way to generate leads—not a shortcut through clinical trials.
The research was funded by the NIH, NSF, an MSU Strategic Partnership Grant, and the Corewell Health-MSU Alliance.