Geneva researchers find metastasis runs on a gradient, not a switch — and built an AI to read it

A University of Geneva team cloned individual cancer cells to discover that tumor spread depends on collective cell behavior, then turned that insight into a prediction tool that outperforms existing staging systems.

Most attempts to predict whether a cancer will spread have chased the same idea: find the gene, the mutation, the single molecular signal that flips a tumor cell from stationary to deadly. Decades of work have produced a long list of candidates and no reliable answer. Metastasis — the process by which cancer cells leave a primary tumor and colonize distant organs — still causes roughly 90 percent of cancer deaths, and oncologists still lack a precise way to tell which patients face the highest risk.

A team at the University of Geneva (UNIGE) has taken a different approach, and the results, published in Cell Reports01606-7), suggest the field may have been looking at the problem wrong. Rather than searching for a single genetic switch, the researchers found that metastatic potential operates on a gradient — more like a dimmer than a light switch — and that the critical unit of analysis is not the individual cell but groups of related cancer cells interacting with each other.

The study was led by Ariel Ruiz i Altaba, a professor of genetic medicine and development at the UNIGE Faculty of Medicine who has spent years studying how developmental signaling pathways — particularly Hedgehog-GLI signaling — get hijacked in cancer. His lab has long argued that cancer is less an anarchic explosion of rogue cells and more a distorted replay of normal development. The MangroveGS work is that thesis taken to its logical conclusion: if cancer follows developmental logic, metastasis should have readable rules.

Cloning cells to watch them move

The technical challenge is deceptively simple to state and fiendishly hard to solve. To understand what makes a cancer cell metastatic, you need to know its complete molecular identity — which genes are active, at what levels — but the analysis destroys the cell. You cannot sequence a cell and then watch it migrate. Ruiz i Altaba's team worked around this by isolating individual cells from two primary colon tumors, growing each into a clone of genetically identical cells, and then splitting each clone: one portion was sequenced, the other was tested for migratory behavior in lab dishes and in mouse models.

They analyzed the expression of several hundred genes across roughly 30 clones, according to the university's press release. What emerged were not discrete categories of "metastatic" and "non-metastatic" cells, but smooth gradients of gene expression that tracked closely with each clone's ability to move and establish secondary tumors. The team named these metastatic potential gradient genes, or MPGGs.

The concept of a gradient matters. Existing staging systems and biomarker tests tend to draw binary lines — stage III or stage IV, positive or negative for a given marker. The Geneva data suggest that metastatic potential is a continuous variable, and that treating it as binary throws away information that matters for patient outcomes.

It takes a village to metastasize

Perhaps the most striking finding is that individual cells do not determine their own metastatic fate. Instead, metastatic potential emerges from what the researchers call "cell-state ensembles" — dynamic collectives of cancer cells that exist in different states (proliferative, dormant, migratory, differentiated) and whose proportions determine the overall threat. A highly metastatic cell can produce daughters spanning the full range of metastatic behavior, the team found through re-cloning experiments.

This is a meaningful departure from the conventional search for metastasis drivers. Most gene-signature tools treat each cell or tumor sample as an independent data point. The UNIGE model treats the ensemble — the community of cell states — as the unit that matters. It is a shift from molecular reductionism to something more ecological.

The team confirmed the functional relevance of their gradient genes using CRISPR-Cas9. When they knocked out CSAG1 in highly metastatic cells, migratory efficiency dropped more than threefold; overexpressing it in moderately metastatic cells nearly doubled their migration. Similar experiments with ATP11C and VGLL1 showed comparable effects — these genes are not bystanders, according to the Cell Reports paper.

MangroveGS: the prediction tool

The biology would be interesting on its own. But the Geneva team went further, building a machine learning model called Mangrove Gene Signatures (MangroveGS) that converts the gradient gene data into clinical predictions. The key design choice, according to co-first author Aravind Srinivasan, a PhD student in the department, is that MangroveGS uses dozens to hundreds of gene signatures simultaneously rather than relying on a handful of markers. This makes the model resistant to the patient-to-patient variability that has undermined many previous biomarker efforts.

After training, the model predicted colon cancer recurrence and metastasis with nearly 80 percent accuracy, outperforming existing methods, according to the Cell Reports paper. In validation cohorts, it achieved a hazard ratio of 10.8 for distinguishing low-risk from high-risk patients — a measure of how sharply the tool separates outcomes between groups, and a number well above what current staging systems deliver. The authors note this figure comes from the paper and has not been independently confirmed.

The signatures also transferred across cancer types. Gene patterns identified in colon cancer proved useful for predicting metastatic risk in stomach, lung, and breast cancers — cancers of epithelial origin that together account for a large share of global cancer mortality.

The distance between a paper and a clinic

MangroveGS is designed for clinical use, at least in principle. Tumor samples collected in hospitals could be RNA-sequenced, and the model would generate a metastasis risk score delivered through an encrypted platform the team calls the Mangrove portal. Ruiz i Altaba told UNIGE's press office that the tool could "prevent the overtreatment of low-risk patients, thereby limiting side effects and unnecessary costs, while intensifying the monitoring and treatment of those at high risk."

That is a compelling vision, but it remains exactly that — a vision. The study validated MangroveGS against existing patient datasets, not in a prospective clinical trial where the tool guides real treatment decisions in real time. The gap between retrospective validation and clinical deployment is wide and littered with promising biomarkers that did not survive the crossing. The team has filed a patent application, which suggests they intend to pursue commercialization, but no timeline for clinical trials or regulatory submissions has been disclosed.

There is also the question of sample size. The foundational biology was built on roughly 30 clones from two colon tumors — a small base from which to derive universal rules about metastasis. The model's strong performance in larger validation cohorts is encouraging, but independent replication by other labs will be essential before anyone should treat the 80 percent accuracy figure as settled.

Why this matters

The field of AI-assisted cancer prognosis is crowded. A recent systematic review of AI tools for predicting colorectal cancer liver metastasis alone cataloged dozens of approaches. What distinguishes MangroveGS is not the machine learning — the ML is relatively straightforward — but the biology underneath it. By starting from clonal analysis and building up to gene gradients and cell-state ensembles, the Geneva team is feeding the model a fundamentally different kind of input than most competitors use.

For the five-year survival rate of metastatic colorectal cancer — fewer than 20 percent according to a JAMA review — any tool that reliably stratifies risk early enough to change treatment decisions would be significant. MangroveGS is not there yet. But the biological framework it rests on, the idea that metastasis is a collective, graded phenomenon rather than a binary genetic event, may prove more durable than the prediction tool itself.

The research was supported by the Swiss National Science Foundation, the Swiss Cancer Research Foundation, and the DIP of the State of Geneva.