Cancer treatment often faces a frustrating guessing game: which tumors will stay put, and which will spread? Researchers at the University of Geneva have built an AI tool called MangroveGS that answers this question with roughly 80% accuracy, outperforming existing methods and working across multiple cancer types.
The research, published in Cell Reports, represents a shift from treating metastasis prediction as a single-gene problem to analyzing the orchestra of hundreds of genes that actually determine whether cancer cells will migrate.
How It Works
Professor Ariel Ruiz i Altaba’s team at UNIGE’s Department of Genetic Medicine and Development took an unusual approach. Instead of looking for a single biomarker, they cultured approximately thirty cell clones from two primary colon tumors and tracked the activity of hundreds of genes simultaneously.
The patterns that emerged weren’t simple. Metastatic potential wasn’t determined by any single cell’s profile, but by how groups of related cancer cells interact with each other. The AI learned to read these collective signatures rather than hunting for individual smoking guns.
MangroveGS exploits dozens, even hundreds of gene signatures, which makes it particularly resistant to individual variations. Where previous tools might fail because one patient’s tumor expresses a key gene differently, MangroveGS looks at the broader pattern and maintains its predictive power.
Real Clinical Application
The tool isn’t just a research exercise. MangroveGS can work directly with tumor samples collected in hospitals. The workflow is straightforward: cells are analyzed, their RNA is sequenced, and a metastasis risk score is generated and shared securely with doctors and patients through an encrypted platform.
The same gene signatures derived from colon cancer also proved useful for predicting metastatic risk in stomach, lung, and breast cancer. This cross-cancer applicability suggests the tool might identify fundamental mechanisms of cancer spread rather than quirks specific to one tumor type.
What This Could Change
Current cancer treatment often defaults to aggressive intervention because the alternative, missing metastatic disease, is worse. Patients whose tumors would never spread still receive chemotherapy, radiation, or extensive surgery. Those whose cancers will spread might not receive aggressive enough treatment early.
An 80% accurate prediction tool could help calibrate these decisions. Low-risk patients might avoid unnecessary treatment side effects. High-risk patients could receive intensified monitoring or more aggressive early intervention.
The tool could also improve clinical trial design by better selecting participants likely to experience the outcomes being studied.
The Fine Print
Eighty percent accuracy sounds impressive, but it means one in five predictions will be wrong. For individual patients, that uncertainty remains significant. A false negative (predicting low risk when cancer actually spreads) could delay critical treatment. A false positive (predicting high risk for a tumor that wouldn’t spread) could lead to unnecessary intervention.
The research was conducted on colon cancer cells in laboratory conditions and mouse models. While the cross-cancer validation is promising, real-world performance across diverse patient populations and clinical settings needs verification.
The tool also requires RNA sequencing of tumor samples, which adds cost and complexity compared to simpler tests. Whether the improved accuracy justifies this in routine clinical practice remains to be demonstrated through larger studies.
Still, an AI that reads hundreds of genes simultaneously to predict cancer behavior represents a meaningful advance over single-marker approaches. As sequencing costs continue to fall, tools like MangroveGS may become practical additions to oncologists’ decision-making toolkit.