DeltaMSI: Artificial Intelligence-based modeling of microsatellite instability scoring by next- generation sequencing

Background: DNA mismatch repair deficiency (dMMR) testing is crucial for detection of microsatellite unstable (MSI) tumors eligible for immunotherapy. MSI is detected by aberrant indel length distributions of microsatellite marker loci, either by operator-dependent visual inspection of PCR-fragment length profiles or by automated scoring using bioinformatic scripts on next-generation sequencing (NGS) data. The former is time consuming and low-throughput while the latter provides simplified binary scoring of a single parameter, typically the number of discrete integrated peaks of the indel distribution. The aim of this study was to leverage the multi-parametric visual scoring of complex indel distribution panels and high-throughput of automated NGS data analysis into a script trained by artificial intelligence and validate it on a large real-world data set of solid tumors. Results: A script (DeltaMSI) was developed that integrates the full complexity of indel distribution (number of peaks, relative distribution and area under the curve). It was trained on the indel distribution of 36 widely used microsatellite marker loci in a dataset of 133 MMR proficient (pMMR) and 46 dMMR tumor samples, taking loss of MLH1/MSH2/PMS2/MSH6 protein expression measured by immunohistochemistry as independent reference technique. After selection of the optimal modelling and marker panel, the two top-performing models per locus (logistic regression and support vector machine) were then integrated into DeltaMSI for combined prediction of MSI status on 28 marker loci at sample level. A bootstrapping method of 50 simulations was used for training and the final model was evaluated on an independent validation set. Diagnostic performance of DeltaMSI was compared to that of mSINGS, a widely used script for NGS-based MSI detection, with the same bootstrapping method. Finally, we tested the robustness of the optimal model on a large set (N=1072) of unselected, consecutive solid tumor samples in a real-world setting and fine-tuned the thresholding for routine clinical use. Conclusions: DeltaMSI achieved higher robustness at equal diagnostic power (AUC=0.950; 95% CI 0.910-0.976) as compared to mSINGS (AUC=0.876;95%CI 0.823-0.918). Its sensitivity of 90% at 98% specificity indicated its clinical potential for high-throughput MSI screening in all tumor types. Clinical Trial Number/IRB: B1172020000040, Ethical Committee, AZ Delta General Hospital