Mutation load estimation model as a predictor of the response to cancer immunotherapy

Abstract

The determination of the mutation load, a total number of nonsynonymous point mutations, by whole-exome sequencing was shown to be useful in predicting the treatment responses to cancer immunotherapy. However, this technique is expensive and time-consuming, which hampers its application in clinical practice. Therefore, the objective of this study was to construct a mutation load estimation model for lung adenocarcinoma, using a small set of genes, as a predictor of the immunotherapy treatment response. Using the somatic mutation data downloaded from The Cancer Genome Atlas (TCGA) database, a computational framework was developed. The estimation model consisted of only 24 genes, used to estimate the mutation load in the independent validation cohort precisely (R² = 0.7626). Additionally, the estimated mutation load can be used to identify the patients with durable clinical benefits, with 85% sensitivity, 93% specificity, and 89% accuracy, indicating that the model can serve as a predictive biomarker for cancer immunotherapy treatment response. Furthermore, our analyses demonstrated the necessity of the cancer-specific models by the constructed melanoma and colorectal models. Since most genes in the lung adenocarcinoma model are not currently included in the sequencing panels, a customized targeted sequencing panel can be designed with the selected model genes to assess the mutation load, instead of whole-exome sequencing or the currently used panel-based methods. Consequently, the cost and time required for the assessment of mutation load may be considerably decreased, which indicates that the presented model is a more cost-effective approach to cancer immunotherapy response prediction in clinical practice.

Fig. 1

Computational framework used during the construction of the lung adenocarcinoma mutation load estimation model

Fig. 2

Performance evaluation of the mutation load estimation model. a Estimated mutation load vs. actual mutation load using the independent validation data (n = 211). b Survival analysis comparing PFS in patients with the high estimated mutation loads (n = 15) with those with the low estimated mutation loads (n = 15). The log-rank test results indicate that the higher estimated mutation load correlates with improved PFS (p = 0.0003). c ROC curve for the classification of DCB/NDB patients using the estimated mutation load. The red point indicates the optimal discrimination threshold 141. AUC = 0.8744. d Immunotherapy response prediction using the estimated mutation load. Gold horizontal line represents the optimal discrimination threshold, 141

Fig. 3

Performance verification using 10,000 random models. a Empirical distribution of R² between the estimated and actual mutation load for 10,000 random models. b ROC curves for the constructed model and 10,000 random models. Blue line, the ROC curve of classifier based on the mutation load estimation model. c Empirical distribution of AUC statistic for 10,000 random models. d Empirical distribution of the classification accuracy for 10,000 random models

Fig. 4

Performance evaluation comparing the actual mutation load with the estimated mutation load using the melanoma model in an independent validation cohort (n = 333)

Fig. 5

Performance evaluation of the immunotherapy treatment response prediction using the melanoma mutation load estimation model. a ROC curve for the classification of clinical benefits using the estimated mutation load in the anti-CTLA-4 treatment patients. Red point, the optimal discrimination threshold 264. AUC = 0.6270. b ROC curve for the anti-PD-1 treatment patients. Red point, the optimal discrimination threshold 206. AUC = 0.5812. c Immunotherapy response prediction using the estimated mutation load for the anti-CTLA-4 treatment patients. Gold horizontal line represents the optimal discrimination threshold 264. The accuracy of the classification is 0.6494. d Immunotherapy response prediction for the anti-PD-1 treatment patients. Gold horizontal line represents the optimal discrimination threshold 206. The accuracy of the classification is 0.6053