Abstract Introduction: It is becoming increasingly apparent that tumor gene expression and spatial H&E profiling can be used to improve the accuracy of DNA-only tumor profiling for clinically relevant use-cases. Future development and deployment of integrative (multi-omic (DNA+RNA) and/or multi-modality (omics + imaging)) tests for treatment and clinical trial profiling hinges on the demonstration of clear robustness and performance gains of prognostic and predictive integrative predictors in a clinical setting. In this work we present several analyses quantifying the utility of integrative testing. We show that a multi-omic test combining tumor expression signatures with DNA profiles substantially improves tumor-of-origin prediction for cancers of unknown primary and also significantly improves MSI detection and resolution. We also show that a simple multi-modality model combining spatial features of tumor infiltrating lymphocytes of tumors with tumor expression is prognostic for MSI-positive patients. Methods: Multi-omicTumor-of-origin: 163 patient sample expression profiles were profiled using the GenMineTOP test. GenMineTOP is a tumor-normal, joint DNA+RNA capture PMDA-approved test in current clinical service in Japan. The samples were clustered using DNA-only, RNA-only, and joint DNA+RNA profiling signatures. MSI prediction: 940 samples from the TCGA COAD, UCEC, and READ cohorts with matched MMR IHC MSI-calling, DNA-NGS-based MSI calling, and RNA-seq data were utilized to train, test and validate an expression-based pan-cancer MSI caller. Additional orthogonal validation was carried out using GenMineTOP multi-omic profiles. Multi-modality Spatial context of TIL and tumor-expression: 158 MSI+ samples from the TCGA-UCEC cohort with matched fresh-frozen H&E 20X images and RNA-seq data were utilized to mine spatial features of tumor-infiltrating lymphocytes and derive gene expression signatures, respectively. A Cox-proportional hazard model for patient survival using gene expression and spatial TIL features was constructed while regressing out tumor stage, gender, age, and ethnicity. Results & Conclusions: Our multi-omic results quantify the benefit of combining DNA and RNA profiles from GenMineTOP to classify tumor-of-origin from unknown primaries reduces classification error by approximately 50%. Additionally, our expression-based MSI status prediction model not only significantly improves NGS DNA (MSI-sensor) based MSI prediction (log-ratio test, pval < 1e-7) but also strongly suggests that nonlinear differential expression drives the improvement in MSI prediction. Our multi-modality results demonstrate that mining the spatial context of endogenous tumor immune response has nontrivial prognostic utility (p < .05). Citation Format: Bojan Losic, Kenji Tatsuno, Vasanth R. Singan, Hiroki Ueda, Aya Shinozaki Ushiku, Vasanthan Jayakumar, Emi Hattori, Toshimitsu Ichijo, Takuma Sasa, Masakazu Akahori, Shinya Hayashi, Esther Hsiao, Stephanie Feupe, Dan Vo, Tiantian Geier, Shuichi Tsutsumi, Tetsuo Ushiku, Kousuke Watanabe, Katsutoshi Oda, Hidenori Kage, Joy Radecki-Crandall, Donavan T. Cheng, Hiroyuki Aburatani. Integrated multi-omic/modal profiling enables improved characterization of oncogenic processes, with impact to clinical actionability in the GenMineTOP tested population [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 6238.
Read full abstract