TOTEM: a multi-cancer detection and localization approach using circulating tumor DNA methylation markers

Abstract

Background: Detection of cancer and identification of tumor origin at an early stage improve the survival and prognosis of patients. Herein, we proposed a plasma cfDNA-based approach called TOTEM to detect and trace the cancer signal origin (CSO) through methylation markers.

Methods: We performed enzymatic conversion-based targeted methylation sequencing on plasma cfDNA samples collected from a clinical cohort of 500 healthy controls and 733 cancer patients with seven types of cancer (breast, colorectum, esophagus, stomach, liver, lung, and pancreas) and randomly divided these samples into a training cohort and a testing cohort. An independent validation cohort of 143 healthy controls, 79 liver cancer patients and 100 stomach cancer patients were recruited to validate the generalizability of our approach.

Results: A total of 57 multi-cancer diagnostic markers and 873 CSO markers were selected for model development. The binary diagnostic model achieved an area under the curve (AUC) of 0.907, 0.908 and 0.868 in the training, testing and independent validation cohorts, respectively. With a training specificity of 98%, the specificities in the testing and independent validation cohorts were 100% and 98.6%, respectively. Overall sensitivity across all cancer stages was 65.5%, 67.3% and 55.9% in the training, testing and independent validation cohorts, respectively. Early-stage (I and II) sensitivity was 50.3% and 45.7% in the training and testing cohorts, respectively. For cancer patients correctly identified by the binary classifier, the top 1 and top 2 CSO accuracies were 77.7% and 86.5% in the testing cohort (n = 148) and 76.0% and 84.0% in the independent validation cohort (n = 100). Notably, performance was maintained with only 21 diagnostic and 214 CSO markers, achieving a training AUC of 0.865, a testing AUC of 0.866, and an integrated top 2 accuracy of 83.1% in the testing cohort.

Conclusions: TOTEM demonstrates promising potential for accurate multi-cancer detection and localization by profiling plasma methylation markers. The real-world clinical performance of our approach needs to be investigated in a much larger prospective cohort.

Keywords: Multi-cancer early detection; Tumor origin; cfDNA methylation.

Fig. 1

Visualization of the multi-cancer diagnostic markers and the CSO markers in cohort samples. t-SNE algorithm was used for dimension reduction. A Clustering of the 500 healthy controls and 733 cancer patients in the training and testing cohorts using 57 multi-cancer diagnostic markers. B Clustering of the 484 true positive cancer samples identified by the diagnostic methylation score model at 98% training specificity using 873 CSO markers

Fig. 2

Methylation scores of healthy individuals and cancer patients stratified by stage. Samples in the training cohort and in the testing cohort were plotted separately. Two-sided Wilcoxon test was performed on the methylation score between healthy individuals and all cancer patients. Spearman's rank coefficient and one-sided p-value of correlation was calculated between the methylation score and the stage (i.e., healthy, stage I, II, III, and IV)

Fig. 3

Performance of the diagnostic methylation score model. A RoC plots of the model, stratified by cancer type. B RoC plots of the model, stratified by stage. C Sensitivity of the model with 98% training specificity for individual cancer types or for multi-cancer, stratified by stage. Error bars indicate the 95% Wilson confidence interval (CI). The number of samples in the training and testing cohort is shown below the stage and separated by a vertical line

Fig. 4

Performance of the CSO ensemble model. Heat maps showing the number and proportion of cancer samples classified into a given class in (A) the training cohort and (B) the testing cohort. Column labels represent actual sample classes and row labels represent predicted classes

Fig. 5

Performance of the diagnostic model and the CSO model in the independent validation cohort. A RoC curves of the diagnostic methylation score model; B Heat map showing the CSO prediction results for STAD validation samples; C Sensitivity, top 1 and top 2 accuracy of TOTEM for STAD validation samples at different stages compared with the training and testing cohorts. Error bars indicate the 95% Wilson CI