Bayesian network enables interpretable and state-of-the-art prediction of immunotherapy responses in cancer patients

Immune checkpoint inhibitors, especially PD-1/PD-L1 blockade, have revolutionized cancer treatment and brought tremendous benefits to patients who otherwise would have had a limited prognosis. Nonetheless, only a small fraction of patients responds to immunotherapy, and the costs and side effects of immune checkpoint inhibitors cannot be ignored. With the advent of machine and deep learning, clinical and genetic data has been used to stratify patient responses to immunotherapy. Unfortunately, these approaches have typically been “black-box” methods that are unable to explain their predictions, thereby hindering their clinical and responsible application. Herein, we developed a “white-box” Bayesian network model that achieves accurate and interpretable predictions of immunotherapy responses against non-small cell lung cancer (NSCLC). This Tree-Augmented naïve Bayes model (TAN) precisely predicted durable clinical benefits and distinguished two clinically significant subgroups with distinct prognoses. Furthermore, Our state-of-the-art white-box TAN approach achieved greater accuracy than previous methods. We hope our model will guide clinicians in selecting NSCLC patients who truly require immunotherapy and expect our approach will be easily applied to other types of cancer. Structured Abstract Background Immune checkpoint inhibitors have revolutionized cancer treatment. Given that only a small fraction of patients responds to immunotherapy, patient stratification is a pressing concern. Unfortunately, the “black-box” nature of most of the proposed stratification methods, and their far from satisfactory accuracy, has hindered their clinical application. Method We developed a “white-box” Bayesian network model, with interpretable architecture, that can accurately predict immunotherapy response against non-small cell lung cancer (NSCLC). We collected clinical and genetic information from several independent studies, and integrated this via the Tree-Augmented naïve Bayes (TAN) approach. Findings This TAN model precisely predicted durable clinical benefit and distinguished two clinically significant subgroups with distinct prognoses, achieving state-of-the-art performance than previous methods. We also verified that TAN succeeded in detecting meaningful interactions between variables from data-driven approach. Moreover, even when data have missing values, TAN successfully predicted their prognosis. Interpretation Our model will guide clinicians in selecting NSCLC patients who genuinely require immunotherapy. We expect this approach to be easily applied to other types of cancer. To accelerate the uptake of personalized medicine via access to accurate and interpretable models, we provide a web application ( https://pred-nsclc-ici-bayesian.shinyapps.io/Bayesian-NSCLC/ ) for use by the researchers and clinicians community. Funding KAKENHI grant from the Japan Society for the Promotion of Science (JSPS) to H.S (21K17856).