Assessment of accuracy in identifying structural variants using third-generation sequencing for breast cancer in the absence of gold standard

Abstract Background: Detection of genomic structural variants (SVs) in breast cancer was of great clinical interest. Numerous methods were developed to call SVs from third-generation sequencing data, but the results and accuracy of various calling methods differ greatly. Thus, assessing the accuracy of these calling methods became important and challenging especially in the absence of gold standards. This study aimed to compare the performance of two SVs calling methods, denoted by T1 and T2 respectively, without gold standards of SVs. Methods: A total of 155,566 structural variants detected by the two calling methods with third-generation sequencing were collected based on 20 breast cancer patients from Peking University People's Hospital. Utilizing a Bayesian approach that could take account of the dependency between calling results, we assessed the accuracy of T1 and T2 in detecting insertion, deletion, duplication, inversion, translocation, and copy number variant, respectively, in terms of sensitivity and specificity. Sensitivity analysis was conducted to evaluate the robustness of the comparison results. Results: The estimated sensitivity and specificity of T2 were higher than that of T1 in detecting different SVs. The estimated sensitivity of T2 in detecting insertion, deletion, and duplication were 0.9438, 0.9440, and 0.9441, respectively, while the sensitivity of T1 were 0.8750, 0.8743, and 0.8843, respectively. Sensitivity analysis showed that the prior information is critical but the comparison results were robust in a certain range. Conclusions: Results showed that T2 significantly outperforms T1 in a general comparison. The proposed model and obtained results are helpful to evaluate and select better SVs calling methods in the absence of gold standards, which in turn will be conducive to identifying new SVs associated with breast cancer or other diseases.