AI model predicts colon cancer metastasis by reading gene patterns in tumor cells

A research team has developed an AI model that can predict whether colon cancer will spread to distant organs by analyzing gene expression patterns inside tumor cells. The finding carries a specific and consequential implication: cancer metastasis is not random. It follows a biological program encoded in the tumor's molecular activity, and that program can be detected before the cancer actually spreads. For patients diagnosed with early-stage colon cancer, a tool that reliably identifies which tumors are on a trajectory toward metastasis could change the entire treatment conversation.

Colon cancer is the third most commonly diagnosed cancer in the United States and the second leading cause of cancer death. The American Cancer Society estimated approximately 153,000 new cases in 2023. Survival rates drop sharply once the disease spreads. The five-year survival rate for localized colon cancer is around 91%. For cases where the cancer has metastasized to distant organs, that figure falls to approximately 13%. The gap between those two numbers is where early prediction has the most to offer.

How the model works and what it actually detects

Gene expression refers to how actively specific genes are being read and translated into proteins within a cell. Tumor cells do not all behave the same way, even within the same patient's cancer. Some cells in a tumor express genes associated with staying in place. Others express patterns linked to mobility, invasion of surrounding tissue, and the ability to survive in the bloodstream long enough to establish new tumors elsewhere. The AI model was trained to distinguish between these expression signatures by processing RNA sequencing data from thousands of tumor samples.

The model identifies combinations of gene activity that collectively signal a high probability of metastatic behavior. No single gene determines the outcome. The prediction comes from recognizing a coordinated pattern across dozens or hundreds of genes simultaneously, which is exactly the kind of task where machine learning outperforms traditional statistical analysis. A pathologist examining a tissue sample under a microscope cannot detect these molecular patterns. The AI model processes data that is invisible to standard clinical assessment.

The biological program behind cancer spread

The concept that metastasis follows a molecular program rather than occurring by chance has significant implications for how researchers think about treating cancer. If spread is random, prevention is largely impossible. If it follows a predictable biological sequence, there are specific points in that sequence where intervention could theoretically interrupt the process before any cells leave the primary tumor site.

The researchers behind this model found that the gene patterns associated with high metastasis risk clustered around processes involved in epithelial-to-mesenchymal transition, a process in which cells lose their normal structural properties and gain the ability to migrate. They also identified elevated expression of genes associated with angiogenesis, the formation of new blood vessels that tumors use to access circulation. These are not new biological concepts, but the model's ability to detect their coordinated activation as a predictive signal in individual patient tumors is what makes the approach clinically useful rather than just scientifically interesting.

What high accuracy actually looks like in practice

The research team validated the model against patient outcome data from multiple independent cohorts. In their primary validation set, the model correctly identified metastatic cases with a sensitivity of 87% and a specificity of 82%. To put that in plain terms: it correctly flagged 87 out of every 100 patients who actually went on to develop metastases, while incorrectly flagging only 18 out of every 100 patients who did not. For a predictive oncology tool, those numbers are clinically meaningful. They are not perfect, but they are substantially better than current staging methods, which rely primarily on tumor size and lymph node involvement and miss a significant share of patients who later develop distant metastases.

Current clinical staging classifies colon cancer as Stage I through Stage IV based on how far it has physically spread at the time of diagnosis. The problem is that staging tells you where the cancer is now, not where it is going. Many patients diagnosed at Stage II, where the tumor has grown through the bowel wall but has not reached lymph nodes, do go on to develop metastases. Under current guidelines, Stage II patients often receive less aggressive adjuvant chemotherapy than Stage III patients. If the AI model can identify which Stage II patients have tumor gene signatures associated with high metastasis risk, those patients could be offered more intensive treatment before any spread has occurred.

Targeted therapies and what the research opens up

The researchers noted that identifying the specific molecular programs driving metastasis also creates potential drug targets. If a cluster of genes driving cell migration is consistently active in high-risk tumors, and if drugs exist that can inhibit those pathways, testing those drugs in patients who the AI model identifies as high-risk becomes a rational clinical trial design. Several compounds targeting epithelial-to-mesenchymal transition and angiogenesis pathways are already in various stages of clinical development for other cancer types, which means the therapeutic pipeline is not starting from scratch.

The model is not yet in clinical use. It was developed and validated using retrospective data, meaning tumor samples from patients whose outcomes were already known. The next step is a prospective clinical trial, where the model would be applied to newly diagnosed patients and treatment decisions would be tested against its predictions over time. The research team has stated their intention to begin prospective validation in 2026, in collaboration with oncology centers in the United States and Europe.