PROJECT SUMMARY/ABSTRACT Autologous CD19-directed chimeric antigen receptor T-cells (CAR-T) have resulted in extraordinary response rates in relapsing and refractory large B-cell lymphoma (LBCL). However, over 60% of CD19-CAR-T recipients will experience disease recurrence or progression. Most of these patients will die from their disease. Mechanisms of CAR-T treatment failure are partially understood and biomarkers informing patient outcomes and management have limited clinical utility. Our central hypothesis is that orthogonal modalities (e.g., clinical, molecular, genomic, and radiomic [quantitative measures from medical images]) complement one another, together providing information on resistance mechanisms and patient outcomes beyond that accessible through any individual modality. We present results suggesting that machine learning is an effective methodology for synthesizing and modeling multiple sources of data together. Cancer cells harness genomic heterogeneity to evade pressure applied by immunotherapies, such as immune checkpoint inhibitors. Our preliminary data also demonstrate that TP53 genomic alterations strongly determine response to CAR-T. Furthermore, using transcriptomic profiling, we found that cancer cellular pathways required for effective transmission of CAR-T cytotoxic signals are distorted in TP53-altered lymphoma. These early findings provide a proof-of-concept for the utility of genomics to inform disease biology and risk after CAR-T. We hypothesize that tumor genetic aberrations in cellular pathways used by CAR-T cells to exert cytotoxicity drive treatment resistance by rendering cancer cells insensitive to CAR-T stimuli and supporting immune escape. In Aim 1, we will use comprehensive genotypic and phenotypic tumor profiling before and after CAR-T to study the role of a priori determined genes and pathways in mediating inherent and acquired treatment resistance. We also hypothesize that orthogonal modalities for patient and tumor profiling are complementary, and their integration into a unified, multimodal model could accurately predict CAR-T outcomes. In Aim 2, we will synthesize data from multiple modalities and use machine learning algorithms to predict CAR-T response and identify novel biomarkers. To meet our goals, we have compiled one of the largest CAR-T patient and sample biobanks. A group of leading experts in immunology, genetics, pathology, radiology, machine learning, and bioinformatics will guide the candidate in this multidisciplinary work. If successful, we expect our combinatorial approach to uncover genetic features underlying inherent and acquired CAR-T resistance and identify new druggable targets. Furthermore, our machine learning approach will support treatment personalization by establishing decision support systems and identifying biomarkers of high-risk patients. Finally, we will introduce novel methodologies for modeling CAR-T outcomes, which are extendable to other forms of treatment.