Presenter(s)
S. Chen1, G. Koushnir2, R. P. Nattamai Malli2, T. Shor2, D. Khankin2, O. Landau2, T. Lederer2, V. Fomin2, A. Weiss2, and N. T. Pfister2,3; 1University of Alabama at Birmingham, Birmingham, AL, 2Numenos, New York, NY, 3Department of Radiation Oncology, The University of Alabama at Birmingham School of Medicine, Birmingham, AL
Purpose/Objective(s): Clinical trials are unable to isolate treatment effects on an individual basis as individual features are lost within comparisons of large cohorts. To retain individual predictive capability within large datasets, we previously utilized continuous average treatment effect (CATE) modeling to identify a continuous measurement of drug benefit on the OAK and POPLAR metastatic non-small cell lung cancer (mNSCLC) clinical trials comparing atezolizumab to docetaxel. The model predicts an immunotherapy (IO) vs chemotherapy (CHT) benefit score, identifying patients likely to benefit more from IO in different cancer types based on pre-treatment tumor gene expression. Radiation therapy (RT) induces transcriptional changes in cancer cells that can be captured as a differential gene expression (DGE) signature. This can be superimposed onto tumor gene expression to simulate the effect of RT for RT-naïve tumors. In this study, we hypothesized that we could synthetically treat patients with RT by applying an RT-induced DGE change to baseline tumor RNA sequencing to assess whether RT synergizes more with IO or CHT across different tumor types.
Materials/Methods: We performed a feasibility study to assess if DGE could be used to simulate radiation treatment synergy, which we assessed would enrich for inflammatory immunotherapy biomarkers for patients predicted to have IO-RT synergy. Robust RNA sequencing data from a human lung cancer cell line (pre-treatment and 72 hours following 4 Gy of radiation) was used to generate the RT DGE signature. The mNSCLC clinical trial CATE model was used to predict benefit scores on each gene set across more than 20 tumor types using TCGA datasets. Pre- and post-RT benefit scores were then calculated for every patient. RT-IO synergy is represented by positive scores, and RT-CHT synergy (i.e., RT reduced benefit to IO) is represented by negative scores. Patients at each extreme were compared to identify biomarkers predicting synergistic and antagonistic impact.
Results: All tumor types demonstrated significantly higher RT-CHT synergy than RT-IO synergy. Small patient subsets were identified in most tumor types with predicted RT-IO synergy, but this could not be distilled into a simple prediction. Squamous cell lung cancer and HPV-negative head and neck cancer had the largest proportion of patients with predicted RT-IO synergy, for which a consistent gene set, enriched for immune-related genes, was identified.
Conclusion: Simulating radiotherapy via DGE signatures can be used as a reverse translational tool to identify patients likely to experience a synergistic effect with addition of RT. Based on features the patients identified, a complex gene expression biomarker predictive of RT-IO synergy could be utilized to stratify patients for combination treatment selection.