AI and Machine Learning Advance Cancer Biomarker Discovery and Treatment Strategies

Researchers led by Xian-Yang Qin at the RIKEN Center for Integrative Medical Sciences in Japan have developed a score that predicts the risk of liver cancer. Published in the scientific journal Proceedings of the National Academy of Sciences, the study establishes that the protein MYCN drives liver tumorigenesis, specifically of the type of tumors found in the deadliest subtype of liver cancer.

Liver cancer, or hepatocellular carcinoma, is the cause of more than 800,000 deaths worldwide every year. The mortality rate is very high because the cancer often remains undetected until the late stages and because the recurrence rate is between 70% and 80%.

The MYCN gene is recognized as a contributor to liver cancer that develops from damaged livers, but exactly how has remained unclear. The researchers reasoned that if its overexpression directly leads to liver tumorigenesis, it would be an ideal candidate as a biomarker and for further study. To test their theory, the team first used a hydrodynamic tail vein injection-based transposon system to insert MYCN (the transposon) into the mouse liver genome.

The team found that when they used the system to overexpress MYCN with always-active AKT, 72% of the mice developed liver tumors within 50 days. A variety of tests showed that these tumors had all the characteristics of human hepatocellular carcinoma. Tumors did not develop when overexpressing one or the other of these genes by themselves.

To characterize the microenvironment, the researchers turned to spatial transcriptomics. This technique shows which genes are turned on in a tissue and exactly where in the tissue that activity is happening. In a mouse model of metabolic dysfunction-associated liver cancer, the researchers used this method to look at gene expression over time and by location as liver tumors developed, focusing on where MYCN was increasing. They discovered a cluster of 167 genes that were differentially expressed in tumor-free sections of liver that had increased levels of MYCN. They named this cluster the "MYCN niche."

Based on the mouse spatial transcriptomics data, the researchers next developed a machine-learning model that can take the characteristics of a given gene-expression pattern and output a score that indicates whether or not it corresponds to a MYCN niche. The model can do this with 93% accuracy.

The MYCN niche score was then calculated for human hepatocellular carcinoma datasets. Patients with higher MYCN niche scores showed a greater risk of tumor recurrence and poorer clinical outcomes. This relationship was stronger when the score was derived from non-tumor tissue than from tumor tissue. The score thus represents a proof-of-concept spatial biomarker that predicts prognosis based on microenvironments that promote tumor formation.

The development comes as artificial intelligence demonstrates broader potential in cancer care beyond diagnostics. Cancer treatment can only target 46 of the 750 known mutations which impact the disease's progression and recurrence. This gap directly aligns with the newly published NHS 10 Year Plan, which calls for earlier diagnosis and a pipeline of affordable, personalised interventions to keep patients cancer-free.

The survivorship gap leaves patients in a vulnerable state post-treatment, often handed the outdated advice to 'watch and wait' for the cancer to return. After surgery or chemotherapy, most cancer survivors, across nearly all cancer types, have no standard maintenance treatment.

Using AI, it's possible to bridge the gap by analysing over 100,000 peer-reviewed research papers to identify low-cost, low-toxicity drugs already on the market that can target the full spectrum of 750 cancer driver mutations. Because these medicines are already generic, they directly support the NHS 10 Year Plan's emphasis on cost-effective innovation that reduces health inequalities.

The power of AI lies in its ability to process vast datasets, including biological, clinical, and pharmacological datasets, to uncover connections humans might miss. For instance, a study in the National Institutes of Health (NIH) database suggests that anti-inflammatory drugs may reduce the risk of disease recurrence in breast cancer patients by 42%. AI can identify combinations of existing drugs that can simultaneously target multiple cancer vulnerabilities, say four or five at once, maximising impact while minimising toxicity.

New therapies often take over a decade and billions of pounds before they're available. Meanwhile, cancer survivors are left without proactive options to prevent recurrence during a time when intervention could dramatically improve outcomes. The survivorship gap, historically underfunded compared to diagnostics and active treatment, represents a significant opportunity to change outcomes through AI-driven drug repurposing and machine-learned biomarker identification.