Detection of circulating tumor DNA without a tumor-informed search using next-generation sequencing is a prognostic biomarker in pancreatic ductal adenocarcinoma

Abstract

The confounding effects of next-generation sequencing (NGS) noise on detection of low frequency circulating tumor DNA (ctDNA) without a priori knowledge of solid tumor mutations has limited the applications of circulating cell-free DNA (ccfDNA) in clinical oncology. Here, we use a 118 gene panel and leverage ccfDNA technical replicates to eliminate NGS-associated errors while also enhancing detection of ctDNA from pancreatic ductal adenocarcinomas (PDACs). Pre-operative ccfDNA and tumor DNA were acquired from 14 patients with PDAC (78.6% stage II-III). Post-operative ccfDNA was also collected from 11 of the patients within 100 days of surgery. ctDNA detection was restricted to variants corresponding to pathogenic mutations in PDAC present in both replicates. PDAC-associated pathogenic mutations were detected in pre-operative ccfDNA in four genes (KRAS, TP53, SMAD4, ALK) from five patients. Of the nine ctDNA variants detected (variant allele frequency: 0.08%-1.59%), five had a corresponding mutation in tumor DNA. Pre-operative detection of ctDNA was associated with shorter survival (312 vs. 826 days; χ²=5.4, P = 0.021). Guiding ctDNA detection in pre-operative ccfDNA based on mutations present in tumor DNA yielded a similar survival analysis. Detection of ctDNA in the post-operative ccfDNA with or without tumor-informed guidance was not associated with outcomes. Therefore, the detection of PDAC-derived ctDNA during a broad and untargeted survey of ccfDNA with NGS may be a valuable, non-invasive, prognostic biomarker to integrate into the clinical assessment and management of patients prior to surgery.

Keywords: Biomarker; Cell-free DNA; Circulating tumor DNA; Next-generation sequencing; Pancreatic ductal adenocarcinoma.

Fig. 1

Error in KRAS at amino acid coding positions G12, G13, and Q61 confounds separation of cases and controls. In (A), the presence of at least one KRAS variant did not discriminate the pre-operative ccfDNA of 14 patients with PDAC (P; cases) from the ccfDNA of four healthy controls (C). No difference was observed between cases and controls using the total number of variants (B) and the Z-score (ie, position-specific error modeling; C) associated with each variant. In (C), note that some Z-score values are low (< 1) indicating that the allele frequency associated with that variant was similar to the noise in the pool of normal controls. Although the difference was not consistently significantly different between cases and controls in both replicates, there was evidence that the number of nonreference allele (NRA) counts associated with a variant may help distinguish signal from noise (D). The solid black arrow in B, C, and d identifies the p.G12D variant shared in both replicates from a control patient. R1 = replicate 1; R2 = replicate 2; solid bars represent the mean value.

Fig. 2

Integration of technical replicates into a multi-step in silico error correction strategy eliminates NGS artifacts. Data is shown for both replicates (R1, R2) for the 14 PDAC patients (P, gray circles) and four healthy controls (C, black triangles). Potential variants common to both replicates are identified as “Shared.” All data is from collapsed consensus reads using UMIs to reduce error. In (A), the baseline number of variants using a Z-score threshold of 2.576 (99^th percentile) to remove error is shown. Identifying variants present in both replicates reduced error in the controls by >83% (B). In (C), the number of potential variants is further reduced by applying a nonreference allele (NRA) count threshold – each identified variant had >3 NRA counts in both replicates. In the controls, this additional criterion further reduced error by >95% (D). Finally, only pathogenic variants in the COSMIC database associated with the pancreas were retained (E). Although errors persisted in individual replicates, 100% of the noise was suppressed by combining replicates (F). In A, C, and E, the inset is a magnification of the ‘Shared’ result. Note, the application of each noise elimination strategy also reduced potential variants in the PDAC patient group which may have removed true positives. R1 = replicate 1; R2 = replicate 2; solid bars represent the mean value.

Fig. 3

Concordance between tumor tissue DNA and pre-operative ctDNA. In (A), the ctDNA variants identified using an unbiased search via integration of technical replicates into in silico error suppression methods were compared to mutations present in the tumor tissue DNA. Three additional mutations were present in ctDNA (asterisks) from four patients that were absent in solid tumor DNA. In (B), using the mutations present in tumor DNA to guide the search for ctDNA improved sensitivity by allowing the replicate data to be combined to increase read-depth along with a reduction in the thresholding criterion because patient-specific variants were sought. PDAC = pancreatic ductal adenocarcinoma

Fig. 4

Survival analysis based on presence/absence of ctDNA in ccfDNA acquired pre- and post-operatively in PDAC patients. In (A), technical replicates were integrated into an error suppression strategy to achieve maximum noise reduction across the entire 118 gene panel to identify ctDNA without using somatic mutations identified from corresponding solid tumor DNA to guide detection in pre-operative ccfDNA. In (B), somatic mutations present in tumor DNA were used to guide detection of ctDNA in pre-operative ccfDNA. Detection of ctDNA in pre-operative ccfDNA with and without a priori knowledge of somatic mutations yielded similar results. Applying similar strategies for detection of ctDNA in post-operative ccfDNA (C and D, respectively) did not identify a significant difference in survival between patients.