Optimizing NGS-based biomarkers for mismatch repair and proofreading deficiency

Testing for mismatch repair and proofreading deficiency (MMRD and PRD) in cancer tissues supports the guidance of immunotherapies and the screening for cancer predisposition syndromes. As more and more patients are receiving molecular tumor diagnostics based on broad sequencing approaches, there is an opportunity to integrate MMRD/PRD testing with the testing for clinically actionable mutations. To reach this aim, there is a need to optimize the NGS-based biomarkers and cutpoints. We analyzed whole exome sequencing (WES) of 1487 patients from the TCGA cohorts of colorectal, stomach, and endometrial cancer (COADREAD, STAD, and UCEC). PCR-based testing and detection of deleterious mutations in POLE or POLD1 served as reference for MMRD and PRD status, respectively. MSI scores were derived from paired tumor-normal (T/N) and from tumor-only (T-only) data using MSIsensor and MSIsensor-pro, cutpoints were optimized by maximizing the balanced accuracy. Missense mutation burden, indel burden, and counts for 96 mutation types were determined from the somatic mutation calls. Using a training-test split (70% vs. 30%), elastic net classifiers were trained and validated for the prediction of MMRD and PRD status. With MSI scores from T/N data a close to perfect separation of mismatch repair deficient from proficient tumors was achieved in the pooled COADREAD/STAD/UCEC cohort (AUC=0.99, optimal cutpoint=2.1%), while MSI scores from T-only data performed less perfect (AUC=0.94, optimal cutpoint=6.7%). The performance of MSI scores from T-only data improved when analyzed separately in COADREAD, STAD, and UCEC which can be explained by differing optimal cutpoints of 12.4%, 3.3%, and 5.6%, respectively. Combining missense and indel burden permitted a close to perfect separation of mismatch repair deficient for proficient tumors. Tumors simultaneously exhibiting MMRD and PRD formed a distinct cluster in UCEC, but not in COADREAD and STAD. Both MMRD and PRD were reflected by characteristic signatures in the partition of the mutations to 96 mutation types. We compared the performance of the following three classifier types to separate tumors with respect to MMRD and PRD: (1) based on MSI scores, (2) combining missense mutation and indel burden, and (3) based on the 96 mutation types. Overall, the three classifiers performed similarly and allowed a good separation in almost all comparisons. Separating proofreading-deficient from -proficient tumors when MMRD was present was the only task that was achieved by none of the classifiers. Simulation studies investigating tumor purity showed that classifiers based on missense mutation/indel burden and mutation types outperformed MSI scores for low tumor purities. The study opens new avenues for MMRD and PRD detection in the setting of clinical WES. We are currently investigating which of the NGS-based markers perform satisfactorily when gathered with large sequencing panels instead of WES.