Abstract 2914: Analyzing the tumor microbiome to predict cancer patient survival and drug response

Abstract

Abstract We investigate whether, for some cancers and drugs, one may generate good predictors of survival and drug response from the tumors' microbial abundances, while controlling for clinical covariates. The input data to our predictors are the normalized set of microbial abundances generated for all samples in the TCGA generated by Poore and colleagues (Poore et al., Nature 2020; 579:567-574). We focused on primary tumors from these data and applied additional filters to further reduce technical variation present in legacy TCGA. We built survival and drug response machine learning (ML) models using this microbial data, while adjusting for clinical covariates, as obtained from TCGA annotations. Survival models were built using Coxnet (Simon et al., J. Stat. Software. 2011; 39, 1-13), employing regularized Cox regression with elastic net penalties. Drug response models were built using a variant of the linear support vector machine recursive feature elimination (SVM-RFE) algorithm (Guyon, et aI., Machine Learning 2002; 46:389-422). For comparison, we also built corresponding models using TCGA gene expression data. We evaluated the predictive performance of our models using standard metrics, employing Harrell's concordance index for measuring the accuracy of patient survival prediction and the area under the receiver operating characteristic for measuring the accuracy of drug response prediction. We find that in four cancer types, adrenocortical carcinoma, cervical squamous cell carcinoma, brain lower grade glioma, and subcutaneous skin melanoma, microbial features were better predictors of survival than clinical covariates alone. However, we found that gene expression is a more powerful predictor of survival, across a wider range of cancer types, than microbial abundances. We find seven cancer-drug pairs where microbiome features are more predictive of patients' response than clinical covariates alone. These seven pairs included chemotherapy treatments for bladder urothelial carcinoma, docetaxel treatment for breast invasive carcinoma and sarcoma, and several treatments for stomach adenocarcinoma. Notably, here we find that microbial abundances are better predictors than expression-based ones, as only five cancer-drug gene expression models performed better than clinical covariates alone. Overall, we find that the tumor microbiome is considerably less predictive than the tumor transcriptome in predicting patient survival, but notably, better in predicting chemotherapy response. Our investigation lays the basis for future research, studying the role of the tumor microbiome (based on abundances derived from sequencing data) in predicting the response to targeted and immunotherapies. Citation Format: Leandro C. Hermida, Edward Michael Gertz, Eytan Ruppin. Analyzing the tumor microbiome to predict cancer patient survival and drug response [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2021; 2021 Apr 10-15 and May 17-21. Philadelphia (PA): AACR; Cancer Res 2021;81(13_Suppl):Abstract nr 2914.