Evaluation of TMB estimates for the prediction of response to immune checkpoint blockage.

Abstract

2632 Background: Tumor mutational burden (TMB) is an emerging biomarker in immuno-oncology (IO). Panel sequencing (PS) is a promising approach for its implementation in clinical practice. Methods: We analyzed TMB as predictor of IO response in a multi-cancer cohort published by Miao et. al. The performance of three large panels (Illumina TSO500, Qiagen TMB [QIAseq] and Oncomine TMB [OTMB]) and two small panels (Illumina TST170 and Oncomine Comprehensive Assay [OCAv3]) was compared to WES by in silico simulations. Separation of responders (PR/CR) from non-responders (PD) was analyzed in the multi-cancer cohort (n = 193), in the lung cancer (n = 36) and in the melanoma (n = 125) subcohort. We also simulated PS in the TCGA pan-cancer cohort. Results: In lung cancer, TMB was strongly predictive for IO response (area under ROC curve [AUC] 0.78-0.94). WES performed (borderline)-significantly better than PS for all five panels (OCAv3: p = 0.011, TST170: p = 0.01, QIAseq: p = 0.048, OTMB: p = 0.063, TSO500: p = 0.11). For the cut-point of 199 mutations, misclassification rates compared to WES (16.7%) were borderline-significantly higher for the small panels OCAv3 (33.3%, p = 0.087) and TST170 (36.1%, p = 0.054), but not for the large panels. In melanoma, TMB was moderately predictive (AUC 0.58-0.63) and WES performed (borderline)-significantly better than the OCAv3, TST170, QIAseq and TSO500 panels. In the multi-cancer cohort, WES did not perform better than PS. TBM estimates from PS include an inherent fuzziness originating from restriction to a limited part of the coding sequence. Based on a random mutation model, we derived a mathematical formula for the coefficient of variation (CV) of TMB: The CV decreases inversely proportional with both the square root of the TMB level and with the square root of the panel size. We showed that the mathematical law is valid for real-word mutation data. Conclusions: Small panels (size < 1 Mpb) performed imprecise in diagnostic TMB estimation. Even using the largest commercially available panels it can be challenging to capture the full predictive information of TMB. The detrimental effect of small panel size can be addressed by using larger panels, but halving the CV of TMB necessitates quadruplication of the panel size.