Plasma-based longitudinal mutation monitoring as a potential predictor of disease progression in subjects with adenocarcinoma in advanced non-small cell lung cancer

This was a retrospective sub-study of adults aged ≥18 years participating in the prospective German Time Series of Biomarkers in Lung Cancer (ZRLK) study. Study participants had not previously been treated systemically for advanced NSCLC, which was diagnosed per the Prevention, Diagnosis, Therapy, and Follow-up of Lung Cancer Interdisciplinary Guidelines from the German Respiratory Society and the German Cancer Society [17]. To be eligible for the sub-study, subjects had to have had histologically proven adenocarcinoma of the lung and at least 2 blood draws during first-line treatment (i.e., prior to clinical disease progression). All subjects provided written informed consent within the context of a prospective observational study, and the study adhered to Good Clinical Practice guidelines and Declaration of Helsinki principles.

Per study protocol, blood samples were collected at baseline, prior to each chemotherapy cycle, and about every 3 months during treatment-free intervals. Twenty mL of peripheral blood was collected in EDTA-coated tubes. Plasma was separated by centrifugation at 1500 x g for 15 min within 2 h of sample collection and stored at − 80 °C until DNA extraction. In total, 445 blood samples from the 71 patient cohort were collected, with minimum 2 and maximum 18 samples per patient. The median number of blood samples per patient was 6. Out of 71 subjects, 47 (66.2%) had a pre-treatment tissue sample available and 42 of the 47 (89.4%) had somatic variants detected. Distribution of baseline characteristics were compared between subjects with tissue biopsies (n = 47) and the remaining subjects (n = 24) by Wilcoxon rank-sum test, Pearson’s chi-square test, and Fisher’s exact test as appropriate.

Cell-free DNA (cfDNA) was extracted from 4 mL of plasma, and tumor DNA was isolated from formalin-fixed paraffin-embedded (FFPE) tumor tissue sections using the AVENIO cfDNA isolation kit (Roche, Branchburg, New Jersey) and KAPA Express Extract kit (Roche, Cape Town, South Africa), respectively. DNA yields were quantified by a Qubit fluorometer (ThermoFisher Scientific, Waltham, Massachusetts). The cfDNA yield had a range of 13 to 1164 ng with the median of 58 ng. The amount of isolated tumor tissue DNA was from 63 ng to 15,420 ng with a median of 1180 ng. cfDNA was sequenced using the AVENIO ctDNA Surveillance assay (Research Use Only; Roche, Branchburg, New Jersey), and tumor tissue DNA was sequenced using a FFPET assay (AVENIO Tumor Tissue Surveillance Kit, Research Use Only; in development; Roche, Branchburg, New Jersey). The median input mass for FFPE samples was 50 ng (range 9–200 ng) and median input mass for plasma samples was 20 ng (range 10–50 ng). High-throughput sequencing was performed on the Illumina NextSeq 500 instrument (Illumina, San Diego, California). Both the AVENIO ctDNA Surveillance Kit and the prototype of AVENIO tumor tissue assays, which are based on Cancer Personalized Profiling by Deep Sequencing (CAPP-Seq) technology [15], utilize molecular barcoding and a hybrid-capture methodology to interrogate selected regions and recurrent mutations from 197 genes commonly mutated in cancer, including 17 important cancer-related genes (e.g., EGFR, KRAS, BRAF, TP53). The average de-duplicated sequencing depth of ctDNA and FFPE tumor tissue DNA was 2779 and 1116, respectively. Somatic variants were identified using the AVENIO ctDNA Analysis Software (Roche, Branchburg, New Jersey), which incorporates integrated digital error suppression [18] to remove polymerase chain reaction duplicates and stereotypical errors.

Somatic variants in cfDNA were derived by filtering out predicted germline variants using an algorithm that is able to detect putative somatic variants with allele frequencies < 30% (not excluding whitelist mutation); this is possible because the algorithm relies on relevant information from multiple somatic mutations rather than on an allele frequency cut-off (described below). After using molecular barcoding and background polishing algorithms (adapted from those described in Newman et al., 2016) [19], the AVENIO ctDNA Surveillance Kit adaptively sets the lower threshold for variant calling based both on the background of a sample and error profile at a given position; no global lower threshold for allele frequencies is applied. Somatic variants in FFPE tumor tissue DNA were derived by post-call filters according to the following criteria: (1) variants had to be in non-repetitive regions of the genome, (2) variants are not predicted to be germline variants (described below), (3) the population frequencies of variants had to be < 1% in the Exome Aggregation Consortium (ExAC) database, and (4) the allele frequencies of variants had to be > 5% (or > 3% if variants were well known hot-spot mutations). Both synonymous and non-synonymous somatic variants were selected for further analysis.

To accurately identify somatic mutations in the absence of matched normal sequences, a machine-learning algorithm (CSMutan) was developed that leveraged mutation databases (ExAC, Single Nucleotide Polymorphism database, 1000 Genomes Project, Catalogue of Somatic Mutations in Cancer, Cancer Genome Atlas) and variant allele frequencies. Using in-house sequencing data of plasma samples whose germline single nucleotide polymorphisms have been identified by comparison with matched normal DNA, we built a classifier to distinguish germline variants from somatic variants with a set of features derived from databases and samples’ variants allele frequency. The detail description of the machine learning algorithm could be found in filed patent application [20]. This classifier was then used to predict germline variant profiles for individual subjects.

The presence (ctDNA – positive) vs absence (ctDNA-negative) of mutant ctDNA was evaluated using a previously described Monte Carlo method [15] indicating whether a given set of variants for each sample was significantly higher than background across the sequenced genomic regions. For a given plasma sample, we started with a set of variants (n), which were high-confidence somatic mutations identified in the matched baseline tissue sample. Using Monte Carlo simulation, we compared the mean allele frequency of the reports against the null distribution of background allele alterations, which most are background sequencing errors, in the plasma sample. In the simulation, we conducted 10,000 iterations, each of which randomly picked n background allele alterations and calculated their mean allele frequency. The empirical p-value of the plasma sample was determined as percentile of mean allele frequency regarding to the null distribution of the background allele alterations. Plasma samples, whose variants with empirical p-values from the Monte Carlo method were less than 0.01, were considered to be ctDNA-positive and otherwise ctDNA-negative.

Subjects underwent routine CT imaging at baseline and after every two therapy cycles (not all data available); if clinically indicated, the frequency of CT imaging was adjusted. CT scans were evaluated by a reference radiologist at the Thoraxklinik, Ruprecht-Karls-University at Heidelberg according to Response Evaluation Criteria in Solid Tumors (RECIST) v1.1 [21]. The relationship between total ctDNA level or the presence/absence of ctDNA mutations and radiological response to treatment was assessed. Total ctDNA level was quantified by summing the allele frequency of each variant detected and normalizing to the total amount of input cfDNA. Progression-free survival (PFS) time was computed as difference in days between date of histological diagnosis and date of diagnosis of a progression of disease event defined by RECIST criteria or date of death of any course. The PFS time was censored for all other subjects at the date of the last imaging assessment. PFS was analyzed descriptively using the Kaplan–Meier method. Individual Cox proportional hazards models were used to assess the significance of each potential adjustment factor (age, sex, ECOG (0 vs. 1 or 2), smoking status (current smoker vs. never smoker vs. quit smoking), and disease stage (IIIA vs. IIIB vs. IV) with PFS. None of these factors were significantly associated with PFS. Two final unadjusted Cox proportional hazard regression models were used to calculate hazard ratios (HRs) with 95% confidence intervals (CIs) for ctDNA mutation status for the first available post-treatment plasma sample and the last available post-treatment plasma sample. Molecular progression was defined as mutational allele frequencies with consecutive increases from baseline.