Biohub grant fuels AI project at HMS linking cells to patient outcomes

A new $1 million grant from Biohub is fueling an effort at Harvard Medical School to develop artificial intelligence tools that link what happens inside individual cells to how patients develop disease and respond to treatment. The work aims to turn vast stores of molecular and clinical data into insights that can guide diagnosis and care.

Leading the project is Marinka Zitnik, PhD, an associate professor of biomedical informatics at HMS and an associate faculty member at the Kempner Institute for the Study of Natural and Artificial Intelligence at Harvard University. Her team is building “virtual cells”—computer models that learn from millions of real measurements of gene and protein activity—and integrating these models with medical records from large patient populations. 

“We can measure what happens inside a single cell in extraordinary detail, but a patient is not a single cell,” Zitnik says. She notes that a disease unfolds over years, across many tissues and many cell types. “Virtual cells let us ask how genes and proteins respond to a drug or mutation,” she says. The challenge is connecting what the virtual cells predict about gene and protein responses to whether an individual patient will benefit from a treatment. “This project is about closing the distance between the scale of a cell and the scale of a patient,” she says.

Foundation models—large AI systems trained on enormous datasets—have revolutionized fields such as language and image recognition. Zitnik’s group is developing foundation models trained on equally rich biomedical data, including single‑cell and spatial transcriptomics, proteomics, and real‑world clinical data from patients with neurodegenerative, cardiometabolic, cancer, and rare diseases.

By bringing these data together, the team hopes to identify molecular biomarkers that support early disease detection, stratify patients into clinically meaningful subgroups, and predict how individuals are likely to respond to specific therapies. 

AI tool improves prediction of cancer immunotherapy response

A central design principle is that the models must be interpretable. Rather than hiding their reasoning as many “black box” AI tools do, Zitnik’s models will highlight biologically meaningful concepts so clinicians and scientists can see why a prediction was made and judge whether to trust it.

To connect these models to real clinical questions, Zitnik’s team will develop software capable of creating a digital representation of a patient, running the most relevant models, and providing risk estimates or treatment options, complete with explanations and uncertainty ranges. The team also plans to establish a shared environment to standardize how models are combined and ensure safety and reproducibility, making it easier for other researchers and institutions to adopt the methods.