# AI Platform GPS for Predicting Gene Expression from Molecular Structures
Scientists from Michigan State University have developed the GPS platform (Gene expression profile Predictor on chemical Structures), which analyzes the structure of chemical compounds and predicts their impact on gene expression. The model operates solely with molecular formulas, without any lab tests. Trained on millions of experimental data points, GPS solves a classification task: up- or down-regulation of gene activity for each compound. This enables virtual screening of drug candidates.
The platform has already identified promising compounds for hepatocellular carcinoma (HCC)—a highly lethal form of liver cancer—and idiopathic pulmonary fibrosis (IPF), where median survival is just three years. Results have been validated through in vivo and ex vivo tests and published in the journal Cell.
GPS Architecture and Training
GPS is trained on a vast dataset of experimental gene expression measurements. The key challenge is noise in biological data. As co-author Jiayu Zhou notes, the model focuses on strong signals while ignoring weak noise. This is achieved through robust learning techniques, where the molecule's structure (SMILES or graph representations) is fed directly into the input.
Process:
- Input data: Molecular structure (formula, graph).
- Processing: Encoding into vector representations (e.g., via graph neural networks).
- Output: Prediction vector for gene expression across thousands of genes (up/down regulation).
- Training: Multi-task classification on millions of compound-gene pairs.
The model requires no prior experiments, accelerating the discovery phase in drug development.
Application to Hepatocellular Carcinoma
For HCC—the third deadliest form of cancer—GPS selected two novel compounds. Testing in mouse models showed a reduction in tumor mass. The compounds modulate key genes involved in hepatocyte proliferation and angiogenesis. This is a first step toward candidates, but further optimization is needed for specificity and bioavailability.
Results for Idiopathic Pulmonary Fibrosis
IPF is a progressive disease with low survival rates. GPS identified:
- One repurposed drug (from existing medications).
- Two novel compounds.
Validation was performed on mice and human lung tissue explants from Corewell Health. The compounds suppress fibrotic pathways (TGF-β signaling), reducing collagen deposition. Senior author Xiaopeng Li highlights the systematic AI approach, in contrast to two decades of traditional failures.
Open-Source Code and Future Prospects
The team from MSU, Stanford, and the University of Michigan has open-sourced the GPS code. The platform can screen any compounds against targets in oncology, fibrosis, and other areas. Supported by NIH and NSF.
Edmund Ellsworth, director of MSU's Center for Medicinal Chemistry, warns: from hits to approved drugs takes years of optimization, preclinical, and clinical trials. Development is an iterative process involving interdisciplinary teams (>20 specialists on the project).
Key Points:
- GPS predicts gene expression solely from structure, no wet lab needed.
- Validated hits for HCC (tumor reduction in mice) and IPF (ex vivo human tissue).
- Open-source code for community-driven screening.
- Scalable to other diseases.
- Noise resistance as the key to accuracy on noisy bio-data.
The platform is shifting the paradigm in computational drug discovery by focusing on prediction-driven hits.
— Editorial Team
No comments yet.