NCI’s Center for Cancer Research (CCR) recently unveiled two new artificial intelligence (AI) models to help predict the patients who are most likely to benefit from immunotherapy. Both patients and physicians can use these AI models, along with some basic clinical information, to gauge response to checkpoint inhibitors.

Most importantly, you don’t have to wonder how the models reach their conclusions—the so-called “black box.” Indeed, rather than a black box that’s hard to decipher, these models rely on “white boxes” with inputs and outputs that are easy to interpret and understand.

The researchers developed two LORIS (or “logistic regression-based immunotherapy-response score”) models for forecasting how patients will respond to checkpoint inhibitors. The first model, a pan-cancer model, is more generalizable and applies across various cancer types. The second model specifically targets patients with advanced, non-small cell, lung cancer (which is one of the world’s largest and most lethal cancer types).

According to the study authors, LORIS outperformed current models in both predicting patient responses and identifying those who were most likely to respond to therapy. Moreover, using the models, the researchers found that even patients previously thought to be poor candidates for checkpoint inhibitors (i.e., those with low tumor mutational burden or PD-L1 expression) could indeed benefit from this therapy. This finding opens the doors of immunotherapy to a much wider population.

In addition, thanks to the white-box nature of LORIS, the researchers were able to gain clues to characteristics that might impact a patient’s response to immunotherapy. According to Dr. Tiangen Chang, the first author of this study, “In both models, we found that prior systemic treatment, such as using chemotherapy as a first-line treatment, could hamper the efficacy of immunotherapy at a later date.”

Dr. Eytan Ruppin from NCI’s CCR and a corresponding author, added, “What we’re learning from LORIS is only the beginning. With advances in large language models, along with comprehensive, multi-omics data, we can expect to gain even more insight into personalizing precision therapy in the future. One of the critical factors to accelerate this process is the sharing of data and code.”