New AI Model Predicts Cancer Spread With Incredible Accuracy

Scientists have developed an AI system that analyzes complex gene-expression signatures to estimate the likelihood that a tumor will spread.

Why do some tumors spread throughout the body while others remain confined to their original location? Scientists still do not fully understand the processes that determine whether cancer cells gain the ability to metastasize. Yet answering this question is essential for improving how patients are treated.

Researchers at the University of Geneva (UNIGE) investigated this problem using cells taken from colon cancers. Their work identified specific factors that influence the likelihood that a tumor will spread. The team also discovered gene expression signatures that help estimate metastatic risk. Using these findings, they developed an artificial intelligence tool called MangroveGS that converts this biological information into predictions for many types of cancer with exceptional reliability. The study, published in Cell Reports, could lead to more personalized care and help scientists uncover new therapeutic targets.

"The origin of cancer is often attributed to 'anarchic cells'," explains Ariel Ruiz i Altaba, professor in the Department of Genetic Medicine and Development at the UNIGE Faculty of Medicine, who led the study. "However, cancer should rather be understood as a distorted form of development."

Genetic and epigenetic changes can reactivate biological programs that were active during the early development of tissues and organs but were later shut down. When these programs become active again in the wrong context, they can drive tumor formation.

In this sense, cancer does not arise randomly but follows an organized biological process. "The challenge is therefore to find the keys to understanding its logic and form. And, in the case of metastases, to identify the characteristics of the cells that will separate from the tumor to create another one elsewhere in the body."

Tracking down metastatic cells

Metastasis is responsible for most cancer deaths, especially in colon, breast, and lung cancers. Today, the earliest detectable sign of metastasis is the presence of circulating tumor cells in the bloodstream or lymphatic system. By the time these cells can be detected, however, they may already have begun spreading through the body.

Scientists have learned a great deal about the genetic mutations that lead to the formation of primary tumors. However, researchers have not identified a single genetic change that explains why some cancer cells leave the original tumor while others remain in place.

"The difficulty lies in being able to determine the complete molecular identity of a cell – an analysis that destroys it – while observing its function, which requires it to remain alive," explains Professor Ruiz i Altaba. "To this end, we isolated, cloned and cultured tumor cells," adds Arwen Conod, senior lecturer in the Department of Genetic Medicine and Development at the UNIGE Faculty of Medicine and co-first author of the study. "These clones were then evaluated in vitro and in a mouse model to observe their ability to migrate through a real biological filter and generate metastases."

The researchers measured the activity of several hundred genes in roughly thirty cloned cells taken from two primary colon tumors. Their analysis revealed clear gene expression gradients that strongly correlated with how easily the cells were able to migrate.

The findings also suggest that metastatic risk cannot be determined by studying a single cell alone. Instead, it depends on the collective interactions among groups of related cancer cells within a tumor.

A highly reliable prediction algorithm

The research team incorporated these gene expression signatures into an artificial intelligence model they developed in Geneva.

"The great novelty of our tool, called 'Mangrove Gene Signatures (MangroveGS)', is that it exploits dozens, even hundreds, of gene signatures. This makes it particularly resistant to individual variations," explains Aravind Srinivasan, PhD student in the Department of Genetic Medicine and Development at the UNIGE Faculty of Medicine and co-first author of the study.

Once trained, the system predicted metastasis and recurrence in colon cancer with nearly 80 percent accuracy, significantly outperforming existing prediction tools. The scientists also discovered that gene signatures identified in colon cancer could help predict metastatic potential in other cancers, including stomach, lung, and breast cancers.

Once trained, the system predicted metastasis and recurrence in colon cancer with nearly 80 percent accuracy, significantly outperforming existing prediction tools. The scientists also discovered that gene signatures identified in colon cancer could help predict metastatic potential in other cancers, including stomach, lung, and breast cancers.

An important step forward for clinical practice and research

MangroveGS could eventually become part of routine clinical care. Doctors would only need a tumor sample. Cells from the sample could be analyzed and their RNA sequenced in the hospital. The system would then generate a metastatic risk score, which could be securely transmitted to oncologists and patients through an encrypted Mangrove portal that processes anonymized data.

"This information will 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," adds Ariel Ruiz i Altaba. "It also offers the possibility of optimising the selection of participants in clinical trials, reducing the number of volunteers required, increasing the statistical power of studies, and providing therapeutic benefits to the patients who need it most."

Reference: "Emergence of high-metastatic potentials and prediction of recurrence and metastasis" by Aravind Srinivasan, Arwen Conod, Yann Tapponnier, Marianna Silvano, Luca Dall'Olio, Céline Delucinge-Vivier, Isabel Borges-Grazina and Ariel Ruiz i Altaba, 29 December 2025, Cell Reports.

DOI: 10.1016/j.celrep.2025.116834