MUIDSI DISSERTATION DEFENSE: Explainable Artificial Intelligence To Stratify Pan-Cancer Patients For Immune Checkpoint Inhibitor Decision Making

Immune checkpoints are a normal part of the immune system. It engages when proteins on the surface of immune cells called T cells recognize and bind to partner proteins on other cells, such as some tumor cells. Immune based therapies such as ICIs work by blocking checkpoint proteins from binding with their partner proteins. This prevents the “off” signal from being sent, allowing the T cells to kill cancer cells. One such drug act against a checkpoint protein called PD-1 or its partner protein PD-L1. Some tumors turn down the T cell response by producing lots of PD-L1. Recent years FDA have granted accelerated approval for the immunotherapy drug on treating specific subgroups or advanced stage of solid tumors. This groundbreaking treatment has shown remarkable promise which could prevent tumors from growing and allowing some patients who receive the treatments to essentially be cured. The fact that some patients treated with immunotherapy have a durable response to cancer shows this treatment’s potential. But despite response rates between 20 and 50 percent in certain groups, such as microsatellite instability-high (MSI-H) colorectal cancer which are characterized by high mutational load, neoepitope formation, and an intense lymphocytic infiltrate when compared to microsatellite stable (MSS) tumors . And some cancers, such as non-small cell lung cancer, rely on some indicators like PD-L1 protein, tumor mutation burden. But how accurate these indicators could predict patient’ responses are still in debating. Moreover, scientists still don’t know why the majority of people with cancer do not respond to immunotherapy drugs. For those patients are not supposed to receive immunotherapy may cause unnecessary long term side effect, such as adrenal insufficiency, and financial burden. Completion of the work we propose here would help us to identify characteristics of patients as subgroups who might benefit from ICIs from multi-omics. In this study, we applied an explainable AI approach for patient stratification. Exploratory mining is contrast pattern-based data mining process. We will integrate these two methods to explore clinically explainable subgroups with phenotypical and genotypical features. These features would be the labels for identify patients who really need ICIs. We applied this method in pan-cancer population and identified the patients’ subgroups based on distinctive features including demographic, phenotypical and genomic characteristics. We believe these distinctive features contribute together to the patients’ response sensitivity to the ICIs. Further wet lab experiments to validate these findings are required prior to initiating clinical trials using these identified features.

Please contact Robert Sanders ( for Zoom information.