Abstract 569: A comprehensive immunoscore to predict immunotherapy responses based on multivariate genomic/transcriptomic features

Abstract

Abstract In recent years, immune checkpoint inhibitors have shown great promise in treating various cancer types; however, only a fraction of patients respond to this type of immunotherapy. Currently, PD-L1 ligand expression and microsatellite-instability (MSI) status are FDA-approved biomarkers to guide selected checkpoint-based immunotherapy; however, due to the complexity of tumor-immune interactions, it is unlikely that any single biomarker will be able to comprehensively predict clinical outcomes across the gamut of immunotherapeutics. Many genomic and cellular features have been shown to contribute to the effectiveness of immunotherapy, including tumor mutational burden (TMB), tumor T cell infiltrate, HLA gene expression, and Treg /myeloid-derived suppressor cell (MDSC) infiltrates. Here we have built a machine learning model trained on multiple features derived from whole exome sequencing (WES) and whole transcriptome sequencing (WTS) data. The resulting patient-specific “immunoscore” describes the likelihood that a patient will respond to a specified immunotherapy. Additionally, we have developed a novel visualization schema, which summarizes the full model immunoscore, as well as the weight and impact of each genomic and transcriptomic feature. In this study, we assessed WES and WTS data from two melanoma cohorts, the first treated with anti-PD1 immunotherapy and the second with anti-CTLA4 immunotherapy. First, we extracted genomic/transcriptomic features around five functional groups: antigen presentation, tumor lymphocyte infiltration, checkpoint gene signatures, interferon-gamma gene signatures, and Treg/MDSC gene signatures. Next, random forest classification was performed to identify significant features and weight the relative importance of each. A final immunoscore was calculated as the patient-specific probability of immunotherapy response, scaled from zero (0% likelihood of response) to ten (100% likelihood of response). We noted that the highest-weighted feature for anti-PD1 response came from the antigen presentation feature group, while the highest-weighted feature for anti-CTLA4 response came from the tumor lymphocyte infiltration feature group, which is consistent with the underlying mechanistic difference between the two checkpoint inhibitors. Finally, our immunoscore has shown significantly better performance compared to any single feature based on 3-fold cross validation (p<0.05). In summary, we have built a machine learning model to predict patient responses to immunotherapies based on WES and WTS data and show that integrating multiple established biomarkers delivers superior performance compared to any individual biomarker. Moreover, our framework can be extended to include novel genomic/transcriptomic features that are identified as mediating immunotherapy response. Citation Format: Mengchi Wang, Aaron Wise, Han Kang, Vitor F. Onuchic, Ali Kuraishy, Sven Bilke, Kristina M. Kruglyak, Shile Zhang. A comprehensive immunoscore to predict immunotherapy responses based on multivariate genomic/transcriptomic features [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2018; 2018 Apr 14-18; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2018;78(13 Suppl):Abstract nr 569.