Background: Tumor mutational burden (TMB) is a promising biomarker for immunotherapy. The challenge of spatial and temporal heterogeneity and high costs weaken its power in clinical routine. The aim of this study is to estimate TMB preoperatively using a volumetric CT-based radiomic signature (rMB).
Methods: Seventy-one patients with resectable lung adenocarcinoma (LUAD) who underwent whole-exome sequencing (WXS) from 2011 to 2014 were enrolled from the institutional biobank of Tianjin Medical University Cancer Institute and Hospital (TMUCIH). Forty-nine LUAD patients with WXS from the Cancer Genome Atlas Program (TCGA) served as the external validation cohort. Computed tomography (CT) volumes were resampled to 1-mm isotropic, semi-automatically segmented, and manually adjusted by two radiologists. A total of 3,108 radiomic features were extracted via PyRadiomics and then harmonized across cohorts by ComBat. Features with inter-segmentation intra-class correlation coefficient (ICC) > 0.8, low collinearity, and significant univariate power were passed to the least absolute shrinkage and selection operator (LASSO)-logistic classifier to discriminate TMB-high/TMB-low at a threshold of 10 mut/Mb. The receiver operating characteristic (ROC) curve analysis and calibration curve were used to determine its efficiency. Shapley values (SHAP) attributed individual predictions to feature contributions. Clinical variables and circulating biomarkers were collected to find potential associations with TMB and rMB.
Results: The top frequently mutated genes significantly differed between the Chinese and TCGA cohorts, with a median TMB of 2.20 and 3.46 mut/Mb and 15 (21.12%) and 9 (18.37%) cases of TMB-high, respectively. After dimensionality reduction, rMB comprised 21 features, which reached an AUC of 0.895 (sensitivity = 0.867, specificity = 0.875, and accuracy = 0.873) in the discovery cohort and 0.878 (sensitivity = 1.0, specificity = 0.825, and accuracy = 0.857 in a consist cutoff) in the validation cohort. rMB of TMB-high patients was significantly higher than rMB of TMB-low patients in both cohorts (p < 0.01). rMB was well-calibrated in the discovery cohort and validation cohort (p = 0.27 and 0.74, respectively). The square-filtered gray-level concurrence matrix (GLCM) correlation was of significant importance in prediction. The proportion of circulating monocytes and the monocyte-to-lymphocyte ratio were associated with TMB, whereas the circulating neutrophils and lymphocyte percentage, original and derived neutrophil-to-lymphocyte ratio, and platelet-to-lymphocyte ratio were associated with rMB.
Conclusion: rMB, an intra-tumor radiomic signature, could predict lung adenocarcinoma patients with higher TMB. Insights from the Shapley values may enhance persuasiveness of the purposed signature for further clinical application. rMB could become a promising tool to triage patients who might benefit from a next-generation sequencing test.
Keywords: immunotherapy biomarker; lung adenocarcinoma; machine learning; radiomics; tumor mutational burden.
Copyright © 2024 Zhang, Yang, Ma, Liu and Ye.