Tumor mutation burden (TMB) is a well-known efficacy predictor for checkpoint inhibitor immunotherapies. Currently, TMB assessment relies on DNA sequencing data. Gene expression profiling by RNA sequencing (RNAseq) is another type of analysis that can inform clinical decision-making and including TMB estimation may strongly benefit this approach, especially for the formalin-fixed, paraffin-embedded (FFPE) tissue samples.
A team led by researchers from the European Organization for Research and Treatment of Cancer (EORTC) for the first time compared TMB levels deduced from whole exome sequencing (WES) and RNAseq profiles of the same FFPE biosamples in single-sample mode. The researchers took TCGA project data with mean sequencing depth 23 million gene-mapped reads (MGMRs) and found 0.46 (Pearson)-0.59 (Spearman) correlation with standard mutation calling pipelines. This was converted into low (<10) and high (>10) TMB per megabase classifier with area under the curve (AUC) 0.757, and application of machine learning increased AUC till 0.854. The researchers then compared 73 experimental pairs of WES and RNAseq profiles with lower (mean 11 MGMRs) and higher (mean 68 MGMRs) RNA sequencing depths. For higher depth, they observed ~1 AUC for the high/low TMB classifier and 0.85 (Pearson)-0.95 (Spearman) correlation with standard mutation calling pipelines. For the lower depth, the AUC was below the high-quality threshold of 0.7. Thus, they concluded that using RNA sequencing of tumor materials from FFPE blocks with enough coverage can afford for high-quality discrimination of tumors with high and low TMB levels in a single-sample mode.
XGBoost binary classifier development workflow
TCGA FFPE samples with matched WES and RNAseq data were reanalyzed to produce callsets in the VCF format. Thirty-two features were introduced to the model in total. RNAseq files along with matching WES files were randomly assigned to two subgroups with variants merged respectively to obtain two sets of variants for each data source. RNAseq variants from the training subgroup were labeled by cross-referencing with the WES callset. Variants matched in WES callset by genomic coordinate were labeled as “signal” and the rest as “noise.” After the model was trained to distinguish between the two classes and validated, variants in the testing subset were reaggregated per sample. Filtering out variants predicted as “noise,” testing per-sample callsets were used to calculate TMB estimates and compared against the respective WES-derived estimates to obtain correlation coefficients and ROC AUC scores.