靶向杂交捕获NGS cfDNA基因panel数据的片段组学分析产生一个正交的基因组维度，用于开发转移性乳腺癌患者的稳健预后模型

Introduction: Analysis of cfDNA fragment distributions, or fragmentomics, yields information about the transcriptional activity of genes but mostly has been applied to WGS and WES datasets. The widespread use of targeted hybrid capture NGS cfDNA gene panels provide a rich data source for mining orthogonal fragmentomic features for developing novel algorithmic signatures for prognosis and predicting response. Here, we applied an uniquely developed fragmentomic computational method for hybrid capture panel data to a prospective study of HER2 negative, metastatic breast cancer patients, who were all treated with paclitaxel and bevacizumab. Experimental Procedures : Fastq files from baseline (n=182) and associated clinical data were downloaded from NCBI (PRJNA745047). Raw sequencing (avg. depth ~2600x) data from a custom panel targeting 48 breast cancer associated genes and 8 cancer gene promoters were processed through a proprietary pipeline to get a fragmentomics matrix comprised of 2936 features. Samples were split into a train-test set (2/3-1/3) balanced on overall survival and key clinical variables. Overall survival (OS) and triple negative (TN) signatures using fragmentomics matrix features in training samples with the glmnet package in R. Association between the OS signature and OS was evaluated using the log-rank test and multivariable cox models. Association between the TN signature and TN status was evaluated using the Wilcoxon test and AUC. Results: An OS signature developed in the training set used 56 fragmentomics features. The association between the signature and OS was directionally uniform in low and high tumor content strata (tumor content defined using the available VAF data) in both training and test data sets (and each nominal log-rank test p-value < 0.05). In the test set, the addition of the fragmentomics signature to a multivariable OS model that already included parameters for TN status, tumor grade, and metastatic site significantly improved the model fit (LRT p-value = 0.009), with a signature hazard ratio of 2.50 (95% CI 1.25-4.99). The TN status signature developed in the training set used 56 fragmentomics features of which 6 overlapped with the OS signature. In the test set, the signature was markedly correlated with TN status (Wilcoxon p-value = 0.0052) with an AUC of 0.75. Summary and Conclusions: This work demonstrates the utility of cfDNA hybrid capture fragmentomic data for developing robust predictors even in low tumor fraction samples of overall survival and TN status. Although additional validation is warranted, these signatures could represent potential biomarkers that could assist in clinical decision making. This work also supports the possibility of retrospectively using any cfDNA panel sequencing dataset for fragmentomics analysis to identify predictive biomarkers.