Deep learning model for predicting immunotherapy response in patients with advanced NSCLC

An innovative deep learning model that uses standard histology images to directly predict an immunotherapy response in patients with advanced non–small cell lung cancer (NSCLC) from a single histology slide holds promise that it can serve as a rich source of predictive biomarkers.

Called Deep-IO, the artificial intelligence model is the first to use supervised training directly based on clinical immunotherapy response outcomes. “This allows the model to learn image-level features that are the most predictive of treatment benefit, rather than relying on proxy markers,” says the lead author of the study, Mehrdad Rakaee, PhD, a senior researcher in the Department of Cancer Genetics at the Institute for Cancer Research at Oslo University Hospital in Norway.¹

The model was developed using 295,581 whole hematoxylin–eosin (H & E)–stained images from 958 patients and provides one of the largest data sets of immunotherapy-treated patients with advanced NSCLC.

Nearly 1000 cases from one participating center in the United States and three centers in the European Union (EU) were validated in a blinded fashion across three independent European centers, explains Dr Rakaee. The US-based development cohort consisted of 614 patients. The model was trained on US data and validated on 344 patients in the EU-based validation cohort with a different imaging format.

The study showed that the deep learning model score independently predicted immunotherapy response in the validation cohort for progression-free survival (hazard ratio [HR], 0.56; CI, 0.42–0.76; p < .001) as well as overall survival (HR, 0.53; CI, 0.39–0.73; p < .001).

The model also was superior to other predictive biomarkers, including the tumor mutation burden, tumor-infiltrating lymphocytes, and PD-L1 in the US cohort and was superior to tumor-infiltrating lymphocytes and comparable to PD-L1 in the validation cohort with a 10% improvement in specificity. When the model was combined with PD-L1 scores, it did better than either biomarker alone in predicting a patient’s response to immunotherapy.

“Despite being trained on US data, the model performed robustly in external validation with consistent accuracy across centers,” says Dr Rakaee.

However, he underscores that the main barrier to implementing the model is that the pipeline currently requires manual annotation of tumor regions and is not fully automated yet. “Our goal is to develop this into a fully end-to-end, user-friendly tool requiring minimal intervention.”

He says that with the rapid progress in computational pathology, especially with the rise of foundation models since 2024, he and his colleagues are currently exploring self-supervised learning approaches.

Although the model was trained specifically for immunotherapy monotherapy, Dr Rakaee says that many patients with advanced NSCLC receive immunotherapy combined with chemotherapy in real-world oncology settings, so he and his colleagues are now working on adapting the model, or similar models, to these combination regimens to ensure relevance across the full range of modern treatment strategies.

For Dr Rakaee, Deep-IO shows that standard H & E slides can serve not only for diagnosis but also as a rich source of response prediction information. “The future of cancer treatment is already stained in H & E; we just need the right tools to read it,” he says.

Commenting on the study, Daniel Morgensztern, MD, an associate professor in the Department of Medicine in the Oncology Division of the Washington University School of Medicine in St. Louis, Missouri, acknowledges the need for better predictors of responses to immune checkpoint inhibitors and that deep learning models, such as Deep-IO, are actively being investigated in both tissues and imaging, “with the latter used for prediction of response to treatment, including immunotherapy, and the probability of malignancy in lung modules.”

He calls the model “intriguing” but says that the performance is still suboptimal and requires refinement. He, as well as the authors of the study, underscores that this is a hypothesis-generating study.

“Furthermore, it is unclear whether it may be applicable for widespread use outside the research institutions,” he says. “Perhaps the greatest use would be to identify the patients who have a higher probability of benefit from immune checkpoint inhibitors in the curative setting and those with metastatic disease who could be treated with single agent immunotherapy, avoiding the use of chemotherapy if the predicted probability of response is very high.”