Can AI Predict Who Responds to Immunotherapy?
Immunotherapy has transformed the treatment landscape for patients with lung cancer, offering the potential for durable responses in some cases. But for many patients, the therapy does not work, and predicting who will benefit remains one of the field’s most pressing challenges.
Now, researchers are turning to artificial intelligence to help close that gap. In a study led by scientists at The University of Texas MD Anderson Cancer Center, a new pathomics approach uses deep learning to uncover patterns hidden within routine pathology slides that may help predict how patients will respond to immunotherapy.
Presented at the American Association for Cancer Research Annual Meeting, the research describes a platform that analyzes tumor tissue to better stratify patients with metastatic non-small cell lung cancer based on their likelihood of benefiting from treatment.
Looking Beyond What the Eye Can See
Pathology slides have long been a cornerstone of cancer diagnosis. While pathologists can identify key features under the microscope, much of the data embedded in these images remains difficult to quantify.
“Immunotherapy has transformed cancer treatment, but only a subset of patients benefit from it, and predicting who will respond remains challenging,” said Rukhmini Bandyopadhyay, PhD, presenting author of the study and a postdoctoral fellow at UT MD Anderson. “While routine pathology slides contain rich information about the tumor and its surrounding environment, the large amount of complex data can be difficult for human experts to fully quantify.”
At @AACR, excited to see how AI is shaping the future of cancer care. Nice work from Rukhmini Bandyopadhyay (@UTMDAnderson) on a deep learning pathomics platform to help predict #immunotherapy response in lung cancer. #AACR26
— Dr. Patrick Hwu (@PatrickHwuMD) April 20, 2026
Read more at https://t.co/HGvqUCZZUs pic.twitter.com/6fmHyTHpgK
That is where pathomics comes in. By applying machine learning to digital pathology images, researchers can extract large-scale data about how tumors are structured and how they interact with their surrounding environment.
“Pathology slides contain an enormous amount of spatial information about tumor cells, immune cells and how they interact, but it’s very difficult for humans to quantify that consistently at scale,” said Ghulam Rasool, PhD, a researcher in Moffitt Cancer Center’s Machine Learning Department. “The challenge is that these images are extremely large and highly variable, so AI models have to be carefully designed to extract true biological signal while remaining robust across different labs and institutions.”
In this study, investigators developed a deep learning framework known as Pathology-driven Immunotherapy Optimization, or Path-IO. The model analyzes routine pathology slides to identify patterns across the tumor microenvironment and classify patients into higher- or lower-risk groups based on their likelihood of responding to immunotherapy.
A More Complete Picture of Immunotherapy Response
Predicting response to immunotherapy has proved particularly challenging because it depends on complex interactions between tumor cells and the immune system.
Ghulam Rasool, PhD
“Immunotherapy response is driven by complex tumor and immune system interactions, and current biomarkers like PD-L1 only capture part of that biology,” Rasool said. “AI can help by analyzing pathology images to more comprehensively quantify the tumor microenvironment and by integrating that with imaging and clinical data to better estimate which patients are most likely to benefit.”
In the study, the Path-IO model was trained and tested across multiple datasets, including nearly 800 patients treated at UT MD Anderson and additional validation cohorts from Mayo Clinic and Gustave Roussy, as well as a phase 3 clinical trial.
The model successfully ranked patients into groups with significantly different outcomes. Patients identified as high risk had more than double the risk of disease progression or death compared with those in the low-risk group.
Notably, the AI-driven approach outperformed PD-L1, the current standard biomarker used to guide immunotherapy decisions in lung cancer.
Strength in Combining Data
The researchers also found that the model’s predictive performance improved when pathology data was combined with radiology imaging and clinical information.
By integrating multiple data sources, the model achieved stronger predictive accuracy than any single biomarker alone, highlighting the value of a more comprehensive view of each patient’s disease
Moving Toward Real-World Impact
While many AI models in oncology have shown promise, relatively few have advanced beyond early stage development.
“What makes this approach notable is that it’s biologically grounded, relies on routine pathology slides and has been validated across multiple cohorts, including a phase 3 clinical trial,” Rasool said. “That moves it beyond a proof-of-concept toward something that could realistically be used in clinical settings, complementing existing biomarkers rather than replacing them.”
Because the model is designed to work with standard pathology slides already collected in routine care, it could potentially be integrated into existing clinical workflows without significant additional cost or infrastructure.
Still, the researchers emphasize that further validation is needed. The current study is retrospective, and future work will focus on prospective testing and incorporating additional molecular data to refine predictions.
