Abstract Introduction: Detection of cancer is its early stages can potentially improve therapeutic effectiveness. Previously, we demonstrated that ELSA-seq, a machine-learning-aided methylation profiling test, can detect and locate multiple cancers with high accuracy in plasma samples. Here, we evaluated the analytical performance of a refined test version of ELSA-seq, including analytical sensitivity, specificity, repeatability/reproducibility, and robustness. Methods: The classification algorithm and cut-offs for the ELSA-seq test were established in the THUNDER study (NCT04820868). Here, we describe the analytical performance using both plasma samples from the THUNDER study and DNA blends from cell lines. Analytical sensitivity was established by defining the limit of detection (LoD) using in-house cfDNA Methylation Reference Standards (MRS) with lowest DNA input for the assay. In brief, fragmented genomic control DNA (NA24385, Coriell Institute) was used as a diluent to contrive human cancer cell lines with defined mixing ratios (tumor fractions), which was further verified by digital droplet PCR (ddPCR). The LoD was determined by the lowest tumor fractions at which accurate DOC and TOO was reported in at least 95% of replicates. As variant allele frequency (VAF) is widely used as a surrogate measurement for tumour fractions, we also used ultra-deep mutation sequencing to attain VAF as an independent piece of evidence. Analytical specificity was assessed by the true negative rate in 120 plasma samples from age-matched healthy donors. A batch-to-batch repeatability/reproducibility study was carried out using 168 clinical samples processed across multiple reagent lots, instruments, and operators. Robustness was evaluated using common interfering substances that potentially could be present in plasma samples such as hemoglobin, bilirubin, triglycerides, and genomic DNA. Testing was performed using 24 cancer and 24 non-cancer samples, with or without interfering substances. Results: At 5ng input mass (approximating from 1ml plasma), the LOD95 was estimated down to 0.05% (tumor fractions) or 0.02% (VAF) among 6 cancer cell lines. 2 false positives were detected in 120 age-matched healthy donor samples, yielding a specificity of 98.3% (95%CI: 93.5-99.7%). All test results were concordant across multiple reagent lots, instruments, and operators (100% repeatability and reproducibility), and high Pearson correlation coefficients (>99%) were observed in the pair-wise comparison. In addition, none of the substances tested interfered with the assay. Conclusions: These results suggest that ELSA-seq is a highly sensitive test for detection of the trace amount of tumor-derived methylation signals in plasma samples. The performance is highly reproducible and robust, which is critical for clinical implementations. Citation Format: Bingsi Li, Jing Su, Guangliang Zhang, Jiayue Xu, Jianlong Peng, Ya Zhou, Fujun Qiu, Shuai Fang, Xiaofang Wen, Guoqiang Wang, Jing Zhao, Hao Wang, Shangli Cai, Zhihong Zhang. Analytical performance of ELSA-seq, a blood-based test for early detection of multiple cancers [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 5116.
Read full abstract