AI can accurately predict immunotherapy response for multiple cancers using routine pathology slides

What if doctors knew whether their patients were going to respond favourably to therapy before treatment ever started?

This could soon be a reality thanks to new research from an MOHCCN-funded team at the Princess Margaret Cancer Consortium (PM2C). In a new study published in npj Precision Oncology, the team demonstrated that an artificial intelligence (AI)-based biomarker tool called ENLIGHT can predict whether a patient will benefit from immunotherapy by analyzing either RNA from the tumour or pathology slides routinely collected during standard care.

This could help quickly match patients to therapies that are more likely to work for them, while allowing patients unlikely to respond to avoid unnecessary side effects and direct them to alternative therapies or clinical trials sooner.

“Immunotherapy has transformed cancer treatment and improved outcomes for many patients, but only a subset of individuals benefit from these therapies, and existing biomarkers used to predict response can be costly, time-consuming, or require additional molecular testing,” said Dr. Changsu (Larry) Park, a medical oncologist at the University Health Network’s Princess Margaret Cancer Centre (UHN-PM) and one of the paper’s lead authors. “ENLIGHT may help change that.”

The researchers found that ENLIGHT performed as well as or better than several commonly used biomarkers, including PD-L1 staining and tumour mutational burden, and did so across different cancer types. The study also showed that ENLIGHT can either use RNA sequencing data (‘transcriptome profiling’) directly or AI-predicted transcriptomes from routine pathology slides, since patterns associated with immunotherapy response are preserved in the AI-predicted profiles. Importantly, the biomarker remained informative throughout treatment, including at the time of disease progression, helping distinguish patients who continued to benefit from immunotherapy from those whose cancers had developed resistance.

“Identifying which patients are most likely to respond to immunotherapy remains one of the biggest challenges in precision oncology,” said Dr. Park. “Our study shows that AI can extract clinically meaningful information directly from routine pathology slides, creating a potentially faster and more accessible way to guide treatment decisions.”

Unlike many existing biomarkers, ENLIGHT does not require additional tissue samples or specialized molecular testing. Instead, it can use digital images of pathology slides that are already generated as part of standard clinical care to predict immunotherapy response. This approach could expand access to predictive testing in healthcare settings where advanced molecular profiling may not be readily available.

“Because pathology slides are routinely collected for virtually every cancer patient, this technology has the potential to be deployed at scale across many different tumour types,” said Dr. Carlos Diego Holanda Lopes, a medical oncologist at UHN-PM and another lead author on the study. “Ultimately, our goal is to help ensure that patients receive the treatments most likely to benefit them while avoiding unnecessary side effects from therapies that may not be effective.”

To create the tool and test its accuracy, the team had to combine and analyze large amounts of data, including pathology imaging, clinical outcomes and genomic data generated through the MOHCCN. With further validation in prospective studies, the team hopes that one day tools like ENLIGHT can be adopted in clinical settings to identify likely responders before treatment begins, bringing precision oncology to more patients in cheaper and more effective ways.

Paper

Longitudinal validation of ENLIGHT, an AI predictor of immunotherapy response and resistance, in pan-cancer cohorts

Authors

Changsu L. Park, Gal Dinstag, Carlos D. Holanda Lopes, Scott Strum, Jeffrey P. Bruce, Omer Tirosh, Saugato Rahman Dhruba, Danh-Tai Hoang, Tuvik Beker, Eldad Shulman, Anna Spreafico, Philippe L. Bedard, Sofia Genta, Albiruni R. Abdul Razak, Eytan Ruppin, Ranit Aharonov, Lillian L. Siu 

Funding

This study was a collaboration between UHN’s Princess Margaret Cancer Centre, the US National Cancer Institute and Pangea Biomed. The work was led through the Marathon of Hope Cancer Centres Network’s Ontario Cancer Consortium, leveraging the transcriptome sequencing done via MOHCCN for Gold Cohort cases to benchmark against the digital pathology-predicted transcriptome profiles.