A new AI system can predict whether a cancer will spread to other organs with nearly 80% accuracy — and it works across multiple cancer types using the same underlying gene signatures.
MangroveGS, developed by researchers at the University of Geneva, analyzes dozens to hundreds of gene expression patterns to assess metastatic risk. Published yesterday in Cell Reports, the tool could help oncologists decide which patients need aggressive treatment and which can be spared unnecessary chemotherapy.
How It Works
The research team, led by Ariel Ruiz i Altaba at UNIGE’s Department of Genetic Medicine and Development, made a discovery that underpins the system: cancer spread isn’t random. Tumors follow predictable biological programs, with specific gene expression patterns signaling their likelihood to metastasize.
MangroveGS analyzes RNA sequencing data from tumor samples to identify these patterns. Unlike previous approaches that looked at individual genes, this system examines the relationships between dozens or hundreds of gene signatures simultaneously. That makes it resistant to the individual variations that trip up simpler models.
The team trained the algorithm on approximately 30 cell clones derived from two primary colon tumors. They evaluated these cells both in laboratory conditions and mouse models to observe which ones actually migrated and formed metastases. The resulting model correctly predicted colon cancer metastasis and recurrence with about 80% accuracy.
Cross-Cancer Applicability
Here’s what makes this interesting: the gene signatures identified in colon cancer also predicted metastatic risk in stomach, lung, and breast cancers. That suggests something fundamental about how cancers evolve the ability to spread — a shared biological program that transcends tumor type.
This cross-applicability could accelerate clinical adoption. Instead of developing separate prediction tools for each cancer, MangroveGS offers a unified approach.
What This Means for Patients
The clinical implications are straightforward. Oncologists could use the tool to:
-
Spare low-risk patients from overtreatment. Current practice often involves aggressive chemotherapy for patients whose cancers wouldn’t have spread anyway. MangroveGS could identify these patients, reducing unnecessary side effects and costs.
-
Intensify monitoring for high-risk cases. Patients flagged as likely to metastasize could receive more aggressive treatment and closer surveillance.
-
Improve clinical trial design. Better patient stratification means smaller trials with more statistical power, potentially accelerating drug development.
The system is designed for real-world use. Tumor samples are analyzed via standard RNA sequencing at hospitals, then results are transmitted through an encrypted portal to oncologists.
The Fine Print
The 80% accuracy figure comes from the research team’s own validation, and independent testing hasn’t been published yet. The training dataset was small — roughly 30 cell clones from two tumors — though the cross-cancer validation provides some reassurance.
The tool predicts risk, not certainty. An 80% accurate model still gets it wrong one time in five, which means some high-risk patients may be undertreated while some low-risk patients receive unnecessary therapy.
The paper doesn’t discuss regulatory status, commercialization timeline, or cost. Hospital adoption would require validation studies, integration with clinical workflows, and likely regulatory approval as a diagnostic device.
Still, the underlying science is solid. If the results replicate in larger studies, MangroveGS represents a meaningful step toward precision oncology — using AI to match treatment intensity to actual risk rather than treating every cancer patient the same way.