Tracking longitudinal genetic changes of circulating tumor DNA (ctDNA) in advanced Lung adenocarcinoma treated with chemotherapy

We conducted a prospective study with 32 consecutive patients recruited who was histologically confirmed advanced or metastatic lung adenocarcinoma between April 2015 and June 2016, they all received platinum based chemotherapy as the first-line treatment in Cancer Hospital, Chinese Academy of Medical Sciences. Baseline blood samples were taken within a week before the first dose of chemotherapy, serial follow-up samples were obtained after each treatment cycle, the detailed clinicopathological information and blood draw time points for each patient were shown in Additional file 1: Table S1 and Additional file 2: Table S2. Clinical information was collected from electronic medical record system or telephone follow-ups. Clinical staging was determined using chest computed tomography, brain magnetic resonance imaging, and 18F-fluorodeoxyglucose positron emission tomography based on the 7th lung cancer TNM classification and staging system. Response to treatment was examined by computed tomography every two cycles and evaluated according to the Response Evaluation Criteria in Solid Tumors (RECIST) 1.1 as complete response (CR), partial response (PR), stable disease (SD), or progressive disease (PD) [13]. The study was approved by the Independent Review Board (IRB) of Cancer Hospital, Chinese Academy of Medical Sciences and BGI Genomics. (CH-BMS-018, BGI-IRB18035).

Whole-exome sequencing was performed on genomic DNAs from two tumors and matched blood samples. The SureSelectXT Human All Exon V5 capture library (Agilent) for 50 Mb of exonic regions was used to capture the exonic DNA. According to the manufacturer’s instructions, we constructed the sequence library with the SureSelectXT Target Enrichment System for Illumina Paired-End Sequencing Library kit (Agilent). Then DNA sequencing of 100-bp paired-end reads were performed using the Illumina HiSeq4000 sequencer.

Sequencing data was filtered with SOAPnuke (v1.5.0; https://​github.​com/​BGI-flexlab/​SOAPnuke) to remove sequencing adapters and low quality reads. We used BWA (v0.5.9; http://​bio-bwa.​sourceforge.​net/​) with default parameters to align high-quality reads to the NCBI human reference genome (hg19). Picard (v1.54; http://​broadinstitute.​github.​io/​picard/​) and Genome Analysis Toolkit (v1.0.6076, GATK IndelRealigner; https://​software.​broadinstitute.​org/​gatk/​) [14] were used to mark duplicates reads and improve alignment accuracy, respectively.

The potential somatic single-nucleotide variants (SNVs) were called by two software based on paired-alignment files (tumor and normal bam).One was MutTect (http://​archive.​broadinstitute.​org/​cancer/​cga/​mutect) using default parameters(–initial_tumor_lod < 4.0 > , Initial LOD threshold for calling tumor variant–tumor_lod < 6.3 > ,LOD threshold for calling tumor variant–normal_lod < 2.2 > ,LOD threshold for calling normal non-germline–dbsnp_normal_lod < 5.5 > ,LOD threshold for calling normal non-variant at dbsnp sites –pir_median_threshold < 10.0 > ,threshold for clustered read position artifact median –pir_mad _threshold < 3.0 > , threshold for clustered read position artifact MAD–max_alt_alleles_in_normal_count < 1>,threshold for maximum alternate allele counts in normal–max_alt_alleles_in_normal_qscore_sum < 20 > ,threshold for maximum alternate allele quality score sum in normal–max_alt_allele _in_normal _fraction < 0.03 > ,threshold for maximum alternate allele fraction in normal–power_constant_qscore < 30 > , Phred scale quality score constant to use in power calculations), and the other was VarScan (https://​github.​com/​dkoboldt/​varscan) with the following parameters: –min-coverage-normal 10 –min-coverage-tumor 14 –min-var-freq 0.001 –somatic-p value 0.05.

We first got the original bam (Bam File A). Considering the fact that the length distribution of cfDNA fragments have a dominant peak at ~ 167 bp, whereas the distribution peak of ctDNA fragments was ~ 133 bp. And there is positive correlation between the proportion of short DNA (below 150 bp) and the amount of tumor DNA in the plasma of cancer cases [15]. We reconstructed the bam file by selecting the reads whose insert size was smaller than 150 bp, designated as remoulded bam (Bam File B). Following the method calling SNVs showed in Materials and methods section ‘Sequencing data processing’ section, we detected potential somatic SNVs based on paired alignment files (tumor and normal bam).

Here we developed a software to filter the initial point mutations by following conditions: (1) In tumor, the mutant site should have ≥ 4 mutant reads with ≥ 1 reads on each strand. In normal, the mutant site can only have ≤ 2 mutant reads with the normal VAF < 0.01 and isn’t within the dbSNP131 database. They are required with at least 40-fold coverage in normal and tumor. The mutations with supported reads ≥ 1 in normal and present in ExAC with AF ≥ 0.001 or with supported reads ≥ 2 in normal but not present in COSMIC database were removed. (2)The number of mutant reads covered this site must be an outlier in a given window around it, and less than 3 smaller insert and deletion in site <=10 bp from a predicted SNV, but 0 string mismatch base pairs in reads located on this mutant site. In addition, mismatching bases can`t enrich in this region which can determine by our software. Here we select 11 bp region around a predicted SNV as a window to analysis. /93) A power which represents the probability of a mutation detected in the plasma was calculated via our software, and those mutations with detected power more than 80 were kept. /94) A posterior probability of a predicted SNV based on TCGA database was also evaluated by our software, and those mutations with detected posterior probability more than 0.8 were kept. /95) Choose the PASS type mutations via the published perl script (https://​github.​com/​ucscCancer/​fpfilter-tool) with the default parameter.

Finally, all the outcomes were checked one by one whether it is caused by nearby misaligned small insertion and deletion events or sequence similarity in the genome, leading to misplacement of reads in the original bam. Then we can get all the mutations with more confidence. All those mutations were annotated by ANNOVAR(http://​annovar.​openbioinformati​cs.​org/​en/​latest/​) [16].

Mutation spectrum analysis was based on the six possible base changes pre- or post- chemotherapy. Then we displayed the proportion of 96 possible mutation types with the R package lwlegopt (https://​github.​com/​BGI-LuoWen/​lwlegopt). To determine the dynamic of mutational processes in plasma, all point mutations pre- chemotherapy were analyzed using MutationalPatterns R package [17] based on the 30 of COSMIC signatures (https://​cancer.​sanger.​ac.​uk/​cosmic/​signatures).

All statistical test were performed in R. In this study, Student’s t-test was used for determining significance of point mutations. Variant allele frequency of mutations was usually tested with a non-parametric Mann–Whitney U test, as well as gene expression between normal and tumor samples, but not for the variant allele frequency of mutations between smoker and nonsmoker which was tested with one-way ANOVA test. In addition, Fisher’s exact test and Cochran-Mantel–Haenszel Chi Squared test were also used in non-silent SNVs, clinical pathological features and Additional file 6: Table S4. And finally, the log-rank test was used to perform overall survival (OS) and progression-free survival (PFS) analysis. For all statistical test used, we assumed that there is independence between data. Box plots showed median values and middle quartile.

In this study, we collected 98 plasma samples from 32 patients during chemotherapy (Fig. 1a). All patients were histologically confirmed advanced (28.1% at stage III and 71.9% at stage IV) lung adenocarcinoma in China (Additional file 1: Table S1 and Additional file 2: Table S2). Of these, 65.6% were men and half of them had malignant pleural effusion. Patients were diagnosed with a median age 60.5 years (range, 43-75 years), and 59.4% of them were smokers. Platinum-based chemotherapy was the first-line treatment for them, 93.7% were treated with pemetrexed combined platinum (53.1% cisplatin and 40.6% carboplatin), and 12 patients received TKI treatment after chemotherapy failure.

To analyze 28 pre-chemotherapy and 68 post-chemotherapy plasma DNAs, we performed a next-generation sequencing (NGS) based on BGI Oseq-ctDNA, a panel that involves 508 cancer-related genes (Fig. 1b; Additional file 3: Table S3).The average coverage of target regions were approximately 274 × for normal control, but 991x for tumors (Additional file 4: Figure S1A). We followed the pipeline to detect somatic mutations (Fig. 1c). Through this pipeline, many candidate point mutations were observed with a mean variant allele frequency (VAF) more than 1% which previous report was counted [18] (Additional file 4: Figure S1B). The genetic alteration landscape of the target coding region across those 27 samples pre- chemotherapy were very similar to TCGA-exome cohort (Fig. 1d).

To further evaluate those point mutations, we performed a screening for two tissue samples pre- chemotherapy which were randomly selected from the 32 patients. Whole exome sequencing was used to analyze these two tissues with an average coverage of 200x. 27 out of 101 (26.7%) alterations were identified in both plasma samples and tissue samples (Additional file 4: Figure S1C, Additional file 6: Table S4).

We identified a total of 1559 point mutations across 98 plasma samples, including 637 non-silents, 262 silents and 660 non-coding mutations. Among the 98 tumor samples, 98% (96/98) carried one or more point mutations during chemotherapy (range, 1–75 mutations), showing a diversity of mutation detection rate of patients during chemotherapy (Additional file 5: Figure S2A). Simultaneously, we found that whole mutation burden and non-silent mutation burden both gradually decreased following platinum-based chemotherapy,with p value 0.034 and 0.013, respectively (Fig. 2a). Moreover, patients with better response to platinum carried less non-silent mutations after chemotherapy compared to those insensitive patients (Fig. 2b). In addition, non-silent SNVs at baseline was higher in patients with response (PR or SDa) compared to non- responders (SD or PD), objective response rate was also greater in patients with higher non-silent SNVs (> median, 7 mutations) compared with low non-silent SNVs (≤ median), although not reaching statistical significance likely owing to small numbers (Additional file 5: Figure S2B).

When we explore the different spectra following chemotherapy, we found that C>G mutations were significantly decreased during platinum treatment, no matter total SNVs (p = 0.039; Fig. 2c) or VAF (p < 0.01; Fig. 2c), especially in patients with objective response (PR or SDa) (p = 0.029; Additional file 5: Figure S2C), suggesting a significant correction between C>G mutations and platinum-based response. Furthermore, C>A transversion, which was reported to be correlated with smoking status [19‐21], the mutation levels were generally low post chemotherapy (p = 0.003; Additional file 5: Figure S2D), especially in non-smokers(Additional file 5: Figure S2E).

We analyzed point mutations of 27 plasma pre-treatment samples, the result indicated that almost 63% (17/27) patients characterized by smoking associated signatures (COSMIC Signatures 4, 16 and 29) [22], of which 64.7% (11/17) were smokers (Fig. 3a), showing that smoking associated signatures were ubiquitous signatures in LUAD [23]. However, there is no relationship between smoking associated signatures and platinum response (Additional file 6: Table S4). Interestingly, we noticed that 2 out of 27 (7.4%) patients displayed a prominent contribution of platinum associated signature (COSMIC Signatures 3) with platinum sensitivity. These findings were further validated in three independent cohorts including TCGA with more than 6.1% cases were characterized by platinum associated signatures (Fig. 3b). The results indicated that mutation spectra and signatures could be factors influencing platinum response [24].

In total, 392 out of 508 (77.2%) cancer-related genes were identified from 98 blood samples (Fig. 4a), including several previously reporter genes: TP53 (pre vs post: 25.9% vs 5.6%),KRAS (pre vs post: 18.5% vs 5.6%), EGFR (pre vs post: 11.1% vs 5.6%), STK11 (pre vs post: 3.7% vs 4.2%), GNAS (pre vs post: 3.7% vs 9.9%). Of these common reporter genes, some non-silent protein-coding mutations were observed both pre- and post- chemotherapy (Fig. 4b), showing the important role of these genes in LUAD. We found that VAFs of some driver mutations decreased following platinum treatment, and some hotspot mutations, such as EGFR(L858R), KRAS (p.G12C), were still present post chemotherapy(Fig. 4c). Besides reporter genes, chromatin modifying gene SETD2 was also commonly mutated (pre vs post: 11.1% vs 7%), as well as NSD1 [25] (pre vs post: 11.1% vs 4.2%), LRRK2 (pre vs post: 11.1% vs 7%) and ALK (pre vs post: 7.4% vs 5.6%). Mutations in genes NOTCH1, NCOA2, MYC, ATRX and RPTOR were significantly increased after platinum chemotherapy in the plasma samples (Additional file 7: Figure S3A), whereas few genes were almost absent after treatment, such as HDAC4, USP9X and POLQ (Additional file 7: Figure S3B), suggesting that neoplastic cells carrying a specific set of somatic mutations were sensitive to platinum treatment. Nonetheless, persistence of genetic mutations like TP53 (p.V41L), MYC (p.Q48H), ALK (p.V467L), ALOX12B (p.I672 M) and CASP8 (p.S301Y) were still observed during platinum treatment (Additional file 7: Figure S3C), showing a link between insensitivity to platinum-based chemotherapy with these base substitutions. Notably, a novel cancer gene ROBO2 (Roundabout Guidance Receptor 2), which plays important roles in apoptosis, motility, angiogenesis and invasion of cancer cells [26, 27], were also mutated in a set of patients (pre vs post: 7.4% vs 1.4%) (Additional file 7: Figure S3D). Moreover, mRNA expression of ROBO2 were significantly decreased in TCGA cohort [28] (p < 0.001). Three independent cohorts showed the variants of ROBO2 (Additional file 8: Table S5). These findings suggest that ROBO2 may play an important role in LUAD tumorigenesis.

To further understand the genetic dynamics process, we examined the distribution of mutations across Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways. 84 point mutations in 30 genes, including EGF/FGF family members and NTRK receptor families, were found in the MAPK pathway (Fig. 5a). Notably, 52% (51/98) samples carried one or more mutations activated the MAPK pathway, implying this pathway would have a pivotal role in LUAD (Fig. 5a). Moreover, we found that many variants in this pathway presented with lower allele frequencies in plasma samples post chemotherapy (p < 0.001; Fig. 5b), suggesting that those somatic mutations in genes within the MAPK pathway would under the influence of platinum-based chemotherapy. We also discovered seven other pathways with VAFs of their mutations significantly changed during platinum treatment, including Notch signaling, TCR/BCR signaling, cell cycle, VEGF signaling, Toll-like receptor signaling and NK cell mediated cytotoxicity (Fig. 5b). In addition, mTOR pathway components were mutated in 10 genes and in more than 16.3% of tumors. Of these mutations, VAFs didn’t have a prominent change (p = 0.25), showing activation of mTOR pathway through somatic mutations contribute less to platinum-combination chemotherapy in advanced LUAD.

91 genomic variations within 47 genes were detected in 56 plasma tumors from 26 patients based on the TARGET database [29] (Additional file 9: Figure S4A). We found that 50% of patients carried potential therapeutic targets both pre and post-chemotherapy, and there exists some potential therapeutic targets presented in 28.1% (9/32) of patients only after chemotherapy. Although some drug mutations were eliminated, there still has a set of genes for which somatic alterations have therapeutic or prognostic implications, such as ALK, EGFR, CDK12 and so on (Additional file 9: Figure S4B). Notably, mutations participating in activating the ERK, PI3K, MTOR pathway were frequent, with 35 alterations in 28 cases occurring in 13 of 17 evaluated genes. Of which, EGFR was the most frequently mutated gene in LUAD (7.14% of tumors), following by ALK (6.12% of tumors) (Additional file 9: Figure S4C). Based on the CIViC database [30], we identified an original EGFR-activating mutation (S768I) in one patient, which can be targeted by Erlotinib and Gefitinib. Simultaneously, KRAS G12C mutation was detected in four patients shown to be actionable by ARS-853, EGFR Inhibitor, Docetaxel, and Selumetinib (AZD6244). Taken together, these data suggest that ctDNA analysis for LUAD patients can yield significant clinical relevance and detect potentially targetable genes in patients, which can provide guidance for individualized therapy during chemotherapy.

Next, we examined the differences of clinical features between chemotherapy sensitive versus insensitive populations, the result demonstrated that more smokers tend to exist in insensitive group (p = 0.074; Fig. 6a). Interestingly, this association between smoking status and treatment efficacy was not found in patients treated with pemetrexed and carboplatin(p = 0.594), but existed in patients who treated with pemetrexed combined with cisplatin(p < 0.05) Above results suggest that smoking may affect the efficacy of platinum chemotherapy, especially pemetrexed combined with cisplatin treatment. Moreover, the survival analysis revealed that patients treated with pemetrexed and carboplatin got longer survival than pemetrexed combined with cisplatin, though with no statistical significance. (Figure 6b).After first line chemotherapy failure, 37.5%(12/32) patients received TKI treatment, the multivariable analysis showed TKI therapy was significantly associated with better outcome, which was significant independent of other factors (Additional file 10: Figure S5).