New Algorithm Identifies Tumors With a Higher Risk of Spreading Throughout The Body

Researchers have developed a new artificial intelligence model capable of predicting the risk of metastasis by analyzing gene activity in tumor cells. To do this, they studied individual cells from colon tumors grown in the laboratory and identified genetic patterns linked to the cancer's ability to spread. These patterns were used to train the MangroveGS system, which demonstrated high accuracy in predicting recurrence and metastasis in various types of cancer. The method can help doctors identify patients at higher risk and better adapt treatment strategies.

One of the biggest questions in medicine is understanding why some tumors behave in a relatively controlled way, while others spread throughout the body and become much more dangerous. Sometimes, two seemingly similar tumors, even in the same patient, can have completely different evolutions.

One of them can be removed surgically and never cause problems again. Another, however, can release cancerous cells into the body even before diagnosis, forming new tumors in distant organs. This process is called metastasis and is responsible for most cancer deaths, especially in colon cancer.

To improve patient treatment, doctors try to predict which tumors are most likely to spread. However, current prediction methods are often based only on tumor size, the appearance of cells under a microscope, or some known genetic alterations.

These strategies do not always reflect the actual behavior of tumor cells. Therefore, scientists are seeking new ways to better understand what makes some cancer cells more capable of migrating and forming new tumors.

In this study, researchers investigated something called metastatic potential, which is the likelihood of a cancer cell leaving its original tumor and starting a new tumor elsewhere in the body. Previous studies had already suggested that not all cells in the same tumor have the same potential to spread. Some seem much more "traveling" than others. This means that even within a single tumor there is a great diversity of cellular behaviors.

To study these differences in detail, the scientists used samples of intestinal tumors taken from patients. From these samples, they isolated individual cells and cultured each one separately in the laboratory.

Each cell gave rise to a small group of genetically related cells, called a clone. Since each clone arose from only one original cell, the researchers were able to compare how different cells from the same tumor behaved.

Next, the scientists analyzed which genes were active in each clone. Genes are instructions present in DNA that guide the functioning of cells. By measuring the activity of thousands of genes at the same time, the researchers were able to observe patterns associated with the cells' potential for spreading.

They discovered a group of genes whose activity gradually increased or decreased as the metastatic potential of the cells increased. These genes functioned as a “biological thermometer” indicating the risk of cancer spreading.

The results also showed that tumor cells can exist in different cellular states, representing distinct modes of functioning within the tumor. These states are not fixed: cells can change from one state to another over time. The researchers identified specific sets of these cellular states that appear to be associated with increased migration and spread capacity of cancer cells.

After identifying these genetic and cellular patterns, the scientists took the next step: creating a tool capable of predicting the risk of metastasis in real patients. To do this, they developed an artificial intelligence model called MangroveGS. This system analyzes sets of genes identified in the study and calculates the probability of a tumor growing back or spreading to other parts of the body.

To test whether the model actually worked, the researchers applied the system to large databases containing genetic information from patients with different types of cancer. The model was able to predict with great accuracy which patients had a higher risk of disease recurrence or metastasis formation. In addition to bowel cancer, the method also showed good results in pancreatic, lung, stomach, bladder, cervical, and possibly breast and prostate cancers.

Overall, the study suggests that many types of cancer may share a similar biological mechanism that controls the ability of cells to spread throughout the body. By identifying this mechanism and translating it into an artificial intelligence model, the researchers created a promising tool to predict cancer behavior and help doctors choose the most appropriate treatments for each patient.

READ MORE:

Emergence of high-metastatic potentials and prediction of recurrence and metastasis 

Aravind Srinivasan, Arwen Conod, Yann Tapponnier, Marianna Silvano, Luca Dall’Olio, Céline Delucinge-Vivier, Isabel Borges-Grazina, and Ariel Ruiz i Altaba

Cell Reports. Volume 45, Issue 1116834January 27, 2026

DOI: 10.1016/j.celrep.2025.116834

Abstract:

What makes a cancer highly metastatic is not known. Here, we inquire on the metastatic potential (MP) of tumor cells, which reflects their probability to emigrate from the primary tumor to new sites to form secondary cancers. We determine the transcriptomic landscapes of single-cell-derived clones in hybrid EMT space and define metastatic potential gradient genes (MPGGs) that linearly track MP strength. Perturbation of selected MPGGs and linked processes reveals a dynamic cellular and molecular framework of what we define as “cell-state ensembles” underlying the emergence of high MPs. To test if MPGGs predict cancer recurrence, we build the MangroveGS machine-learning model with “gene signature ensembles”: MangroveGSMPGGs robustly predicts patient tumor recurrence and metastases, outperforms all other signatures and staging systems tested, and can be extended to multiple cancer types of epithelial nature. Our findings uncover an unsuspected shared strategy for the onset of metastases that underlies clinical outcome.