Serial circulating tumor DNA identification associated with the efficacy and prognosis of neoadjuvant chemotherapy in breast cancer

Circulating tumor DNA (ctDNA) provides a promising noninvasive alternative to evaluate the efficacy of neoadjuvant chemotherapy (NCT) in breast cancer. Herein, we collected 63 tissue (aspiration biopsies and resected tissues) and 206 blood samples (baseline, during chemotherapy (Chemo), after chemotherapy (Post-Chemo), after operation (Post-Op), during follow-up) from 32 patients, and preformed targeted deep sequencing with a customed 1021-gene panel. As the results, TP53 (43.8%) and PIK3CA (40.6%) were the most common mutant genes in the primary tumors. At least one tumor-derived mutation was detected in the following number of blood samples: 21, baseline; 3, Chemo; 9, Post-Chemo; and 5, Post-Op. Four patients with pathologic complete response had no tissue mutation in Chemo and Post-Chemo blood. Compared to patients with mutation-positive Chemo or Post-Chemo blood, the counterparts showed a superior primary tumor decrease (median, 86.5% versus 54.6%) and lymph involvement (median, 1 versus 3.5). All five patients with mutation-positive Post-Op developed distant metastases during follow-up, and the sensitivity of detecting clinically relapsed patients was 71.4% (5/7). The median DFS was 9.8 months for patients with mutation-positive Post-Op but not reached for the others (HR 23.53; 95% CI, 1.904–290.9; p < 0.0001). Our study shows that sequential monitoring of blood ctDNA was an effective method for evaluating NCT efficacy and patient recurrence. Integrating ctDNA profiling into the management of LABC patients might improve clinical outcome. This prospective study recruited LABC patients at Peking Union Medical College Hospital (ClinicalTrials.gov Identifier: NCT02797652).


Introduction
Breast cancer is one of the most common malignancies for women. According to 2018 global cancer statistics, over 2 million patients were diagnosed as having breast cancer and 630,000 died of the disease worldwide [1]. Currently, neoadjuvant chemotherapy (NCT) is a useful treatment for locally advanced breast cancer (LABC) to downstage and render inoperable disease operable [2]. In addition, NCT can reduce the spatial size of tumors too large for a lumpectomy, thus providing an opportunity to conserve the breast and greatly improve the post-operative quality of life of patients [2]. However, the efficacy evaluation of NCT has not yet been well developed. The clinical imaging methods, such as ultrasound and magnetic resonance imaging, are not always in Yidong Zhou, Yaping Xu and Changjun Wang contributed equally to this work. accordance with pathological assessment [3]. Furthermore, conventional tumor markers, such as CEA and CA15-3, have not shown a strong correlation with response to NCT [4][5][6]. Because distant relapse is the major cause of cancer death [7,8], more effective diagnostic tools that stratify patients by distant recurrence risk would be of great clinical significance to deliver personalized healthcare. Overall, the development of novel markers with high prognostic and predictive performance for post-operative relapse risk are crucial to further facilitate clinical decision-making.
Circulating tumor DNA (ctDNA) is derived from the apoptosis, necrosis and secretion of tumor cells and can harbor tumor-specific genetic or epigenetic alterations [9,10]. CtDNA can be detected in various types of solid tumors, including breast cancer [11]. Several studies have reported that mutant ctDNA is correlated with the response to systemic treatments in breast cancer [3,[12][13][14][15][16]. However, most of these studies only focused on limited epigenetic or genetic markers, such as RASSF1A methylation [3] and TP53 variants [13], and breast cancer is a highly genetically heterogeneous malignancy [17][18][19][20]. Such strategies restricted the population that can potentially benefit from ctDNA profiling. In addition, tracking of mutant ctDNA post-operatively could be a useful method for predicting and monitoring distant relapse in breast cancer [12].
In this study, we enrolled 32 of patients with locally advanced and metastatic breast cancer who were eligible for NCT and underwent standardized clinical managements. We collected tumor tissues to profile patient-specific mutation pools using Geneplus-1021 genes targeted sequencing with a well-designed capture panel and traced them in sequential blood ctDNA around NCT and during post-operative followup. Through these efforts, we aimed to explore the utility of ctDNA to evaluate the efficacy and prognosis of neoadjuvant chemotherapy for breast cancer.

Experimental cohort and genomic profiling of tissue
An overview of patient enrollment and study design is shown in Fig. S1. We initially enrolled 65 patients with locally advanced or metastatic breast cancer between May 2016 and March 2017. After further clinical assessment, 23 patients were excluded because they were ineligible for neoadjuvant chemotherapy. In addition, one patient was excluded due to insufficient tumor tissue for genomic analysis and seven patients were excluded due to drop-out during follow-up. Ultimately, 32 patients were included in the study, and 63 tissue and 206 blood samples were collected (Fig. S2). Baselinen (pre-treatment) patient characteristics and clinical staging information are shown in  (Fig. S3). To eliminate the temporal bias in the evaluation of plasma, we assessed the union set of tissue biopsy and resected tissue mutations in the same patient. As a result, at least one mutation was identified in each patient. In total, 151 mutations were detected with a median of 3 (range, 1-23). TP53 (n = 14, 43.8%) and PIK3CA (n = 13, 40.6%) were the most recurrent mutant genes as expected (Fig. 1a). We also evaluated the concordance between ERBB2 amplification at baseline and HER2 status determined by immunohistochemical staining and fluorescence in situ hybridization (Fig. 1b). Among 11 patients with HER2-positive tumors, 10 (90.9%) carried ERBB2 amplification (copy number > 3). All 21 patients with negative HER2 exhibited a normal copy number of ERBB2. The overall concordance was 96.9%. Even for the inconsistent one, the copy number of ERBB2 (2.6) was higher than that of patients with HER2-negative tumors. It did not reach the cut-off value possibly due to the insufficient proportion of tumor cells in tissue sample.

Clinical characteristics and baseline ctDNA
Based on the aforementioned tissue analysis, we tracked so-called tumor-derived mutations in blood ctDNA, including SNVs, Indels and CNV of ERBB2. As a result, at least one tumor-derived mutation could be detected at baseline in 21/30 patients (70%), during chemotherapy in 3/32 patients (9.4%), post-chemotherapy in 9/31 patients (29.0%), and post-operatively in 5/32 patients (15.6%) ( Table 1). All patients with negative ctDNA at baseline demonstrated no mutation in blood ctDNA at any other time point, indicating the importance of determining the baseline status of ctDNA ( Table 1).
At baseline, premenopausal patients had significantly more positive ctDNA than menopausal patients (p = 0.013, Fig. 2a). We also evaluated the variant allele frequency (VAF) for each plasma sample to reflect the circulating tumor burden. For plasma samples with more than one mutation, the VAF was identified as the maximal one. Similarly, the premenopausal cohort showed a higher VAF than the menopausal cohort (p = 0.007, Fig. 2a). In addition, patients with more lymph involvement had a higher proportion of positive ctDNA than those with less lymph involvement (p = 0.011, Fig. 2b). Although not reaching statistical significance, a higher VAF was observed in patients with more lymph involvement (p = 0.073, Fig. 2b). Higher fractions of positive ctDNA (p = 0.007) and VAF (p = 0.043) were identified in patients with advanced stage than in those with early stage tumors (Fig. 2c). Furthermore, the fraction of patients with positive ctDNA associated with high recurrence risk factors, including poor differentiation, hormone receptor negative, HER2 positive, high Ki67, and metastasis (Fig.  S4). The VAF was also relatively high for patients with positive progesterone receptor and metastasis (Fig. S4). Taken together, these results indicate that the baseline ctDNA status is associated with multiple clinical factors; however, they should be further validated in large-scale cohorts.

The peri-neoadjuvant ctDNA associated with the therapeutic efficacy
To explore the variation of cfDNA concentration and ctDNA VAF at different sampling time points, 30 patients with sufficient DNA from blood at baseline, during chemotherapy, and after chemotherapy were analyzed. At baseline, the cfDNA concentration was the lowest (p < 0.001, Fig. 2d), whereas the VAF was the highest (p < 0.001, Fig. 2d). For the dynamic change of ctDNA, four types of patients were present: (1) ctDNA was negative at baseline, during chemotherapy, and after chemotherapy (n = 9); (2) ctDNA was positive at baseline, but negative during and after chemotherapy  positive at baseline and after chemotherapy, but was negative during chemotherapy (n = 7, Fig. S5C). For type 3 and 4, the ctDNA residue during and after chemotherapy may be a clue about the refractory disease and be associated with therapeutic efficacy. Whereafter, we analyzed the correlation between ctDNA status and therapeutic efficacy. Miller-Payne classification was used to assess the pathological response of carcinoma. For N stage (ypN) after NCT, one patient was excluded because the post-chemotherapy blood sample was unavailable. Overall, 21 patients (67.7%) exhibited negative ctDNA during and after chemotherapy, including four pCR patients (Fig. 3). The median decrease in primary tumor volume was 54.6% (range, 17.1-98.0%) and 90.2% (range, 7.7-100%) for patients with positive and negative ctDNA, respectively (Fig. 3). With respect to Miller-Payne classification, grade 4-5 was found in only 10% (1/10) of patients with positive ctDNA but over half (52.38%, 11/21) of patients with negative ctDNA (p = 0.046; Fig. 3, Table S2). Furthermore, we found the for the comparison of ctDNA VAF, the p-value was calculated by Mann-Whitney test. d Change of cfDNA concentration and ctDNA VAF at different time points. The p-value was calculated by one-way ANOVA. p < 0.05 indicated statistical significance median number of involved lymph nodes was 3.5 (range, 0-6) for patients with positive ctDNA and 1 (range, 0-9) for patients with negative ctDNA. Over 10 regional lymph nodes were involved (ypN3) in four patients with positive ctDNA. Nevertheless, no negative patients demonstrated stage ypN3 (p = 0.0067; Fig. 3, Table S2). To determine whether baseline ctDNA status is related to mutational detection of serial ctDNA based on the aforementioned results, we next performed the same analysis in 22 patients with positive baseline ctDNA and drew similar conclusions as expected (for Miller-Payne grade, p = 0.0244; for stage ypN, p = 0.0260; Table S2). These results indicated that peri-neoadjuvant ctDNA status before surgery could predict the pathological decrease in carcinoma in situ, as well as the involvement of regional lymph nodes evaluated post-operatively.

Post-operative ctDNA status related to recurrence
The median follow-up was 23.2 months (range, 16-26.3 months). All patients underwent post-operative adjuvant agents, including radiotherapy (n = 32), aromatase inhibitor (n = 21), and trastuzumab (n = 9). During followup, seven patients (21.9%) experienced distant recurrence, including five LABC and two MBC. We evaluated the prognostic discrepancy between patients grouped by different pathological factors, CEA/CA15-3 level, and serial ctDNA status. As a result, the recurrence rate and DFS performed similarly between patients with different Luminal isoforms, Miller-Payne grade, ypN stage, pathologic response, CEA/ CA15-3 level, as well as cycle 3, and post-chemotherapy ctDNA status (Fig. S6). Although not demonstrating statistical significance, patients with clinical stage IV or positive baseline ctDNA tended to present poorer DFS compared with their counterparts (Fig. S6). Dramatically, patients with positive ctDNA after surgery had a higher risk of recurrence and worse DFS than those with negative ctDNA (Fig. 4a)

Sequential ctDNA analysis in tumor surveillance
For seven patients with distant recurrence, blood samples were collected 3-5 times post-operatively and during followup. We used the VAF to depict the dynamic change of tumor burden for each patient. The CEA/CA15-3 levels at some time points were also taken into consideration. As shown in Fig. 5, the quantitative VAF of ctDNA fluctuated with the clinical performance of each patient. All patients experienced a dramatic increase in ctDNA VAF at relapse, and multiple ctDNA samples were positive between surgery and relapse, except for those of P25. P03 and P08 underwent adjuvant radiotherapy and a sequential aromatase inhibitor regimen, and ctDNA samples collected post-operatively and at recurrence were both positive. The VAF increased despite therapeutic agents and persisted in subsequent time points. Particularly for P03, a blood sample was collected at the end of chemotherapy after recurrence. The therapeutic evaluation was partial response, and a significant decrease in the VAF was observed. For P19, the VAF was stable during adjuvant radiotherapy but increased dramatically at recurrence. The therapeutic evaluation of second-line radiotherapy was PR, and the tumor progressed 112 days later, both of which were reflected in the dynamic change in ctDNA VAF. For P24 and P39, two blood samples were collected during adjuvant radiotherapy and sequential trastuzumab intake, the latter of which was detected with tumor-derived mutations. One hundred and sixty days later, both patients experienced bone or chest metastasis, and ctDNA VAF also increased to a relatively high level. For P24, the ctDNA burden then decreased significantly due to subsequent chemotherapy and lapatinib intake. For P25, only blood collected at recurrence was detectable with tumor-derived mutations. P41 underwent adjuvant radiotherapy and a sequential aromatase inhibitor regimen with trastuzumab. CtDNA VAF was initially stable and then increased when chest, lymph node, and bone metastasis arose. In consideration of the hypothesis that positive ctDNA represents the post-operative minimal residual disease before recurrence, tumor relapse could be predicted in advance with a median interval of 6.9 months (range, 0-12.1 months). However, the sensitivity of conventional CEA/CA15-3 in tumor surveillance was unsatisfactory. Only one (P25) patients exhibited elevated CEA/CA15-3 post-operatively; two (P08, P25) had elevated CEA/CA15-3 at recurrence (Fig. 5). Furthermore, P24 and P25 showed elevated CEA/ CA15-3 levels at 6.9 and 7.3 months before recurrence, respectively, indicating an advantage in relapse prediction over ctDNA (3.3 and 0 months) (Fig. 5). Altogether, the combination of CEA/CA15-3 and ctDNA may improve treatment monitoring in patients with locally advanced or metastatic breast cancer.

Discussion
Here, we report a study about the application of ctDNA to monitor neoadjuvant chemotherapy for breast cancer. We enrolled dozens of patients with locally advanced or metastatic breast cancer and collected serial blood samples around neoadjuvant chemotherapy, after surgery, and during follow-up. In addition, tissue biopsies were collected at diagnosis and during surgery to fully reveal the patient-specific mutational landscape and traced these tumor-derived mutations in serial blood ctDNA. As a result, peri-neoadjuvant ctDNA status was related to Miller-Payne classification and ypN stage determined through post-operative pathologic evaluation, and post-operative ctDNA status was associated with disease recurrence. Furthermore, the dynamic mutational burden of ctDNA was correlated with the clinical outcomes of recurrent patients, and distant recurrence could be predicted months in advance via the quantification of tumor-specific mutations in ctDNA. Compared with a single marker, the combination of CEA/CA15-3 and ctDNA seem to be a more promising and sensitive method in tumor surveillance.
To date, the prognosis of breast cancer has been significantly improved by multiple new agents. Based on current guidelines, neoadjuvant chemotherapy is the routine treatment for locally advanced breast cancer. This therapy has been increasingly used to shrink tumors enough to qualify for breast-conserving surgery and assess early in-vivo response to systemic therapies [21,22]. The determination of pCR after surgery is the main surrogate marker for adjuvant agents and long-term survival [23,24]. Nevertheless, this correlation mainly exists at an individual patient level and is not consistent at the population level. One published study failed to show a relationship between the improved pCR rate and improved DFS [25]. Some aggressive phenotypes, such as triple-negative breast cancer, display high pCR rates but also have frequently early metastatic relapses [26]. Based on aforementioned issues, two areas of research are critical to improving care for patients with breast cancer undergoing neoadjuvant chemotherapy: (1) new markers are urgently needed to predict clinical response to neoadjuvant chemotherapy to better guide treatment planning and management; and (2) in addition to pathological evaluation, how we can determine better prognostic indicator.
To solve these questions, we focused on ctDNA, a novel and tumor-specific biomarker that has been broadly explored in recent years. In breast cancer, ctDNA detection is associated with clinical stage [11]. For patients with stage III and IV, the detectable rate is much higher than those with stage I and II. In this study, most patients were stage III, accompanied by several patients at stage IIb and IV. The detectable rate of 70% at baseline is in concordance with the previous study [11]. Patients with advanced clinical N and M stages demonstrated a higher rate than those with early stages; however, it did not reach statistical significance, perhaps due to the small sample size. In addition, we found that premenopausal patients were more likely to have positive ctDNA than older patients, which may support the notion that premenopausal breast cancer is more aggressive than post-menopausal breast cancer [26,27]. Baseline ctDNA status was strongly associated with the detectability of serial ctDNA. In patients with negative ctDNA at baseline, no tumor-specific ctDNA detected in any blood samples. This finding suggests that the detectability of ctDNA is an inherent factor for each tumor. Therefore, determining a patient's ctDNA status at baseline is most important prior to other ctDNA-related exploration. This finding may help enhancing the sensitivity of serial ctDNA test, as well as avoiding excessive detection with higher efficiency and less medical expense.
The vast majority of patients with positive ctDNA at baseline demonstrated an obvious decline in the VAFindicating substantial downstaging-after neoadjuvant chemotherapy. However, for some patients, tumor-derived mutations were still detectable in blood samples at cycle 3 or post-chemotherapy. When integrated with pathological evaluation, we found that these patients have a worse prognosis than those whose tumor-specific ctDNA disappeared. The significant correlation between peri-neoadjuvant ctDNA status and pathological evaluation suggests that serial ctDNA analysis around neoadjuvant treatment may be used to guide subsequent therapeutic strategies, such as escalation or de-escalation of systemic treatment. Nevertheless, the criteria of differentiation should be further explored with larger-scale studies.
Compared with baseline, ctDNA remained detectable in much fewer patients after chemotherapy (29%) and after surgery (15.6%), suggesting that ctDNA detected at baseline is mostly derived from the primary rather than the distant metastatic sites. This may explain why baseline ctDNA status is not prognostic. Namely, ctDNA can reflect minimal residual disease, which is associated with the risk of relapse, only when the primary tumor has been removed. Post-chemotherapy ctDNA is also not prognostic possibly due to the incomplete elimination of markers derived from primary sites, which can be modestly drawn from the relatively higher positive rate than post-operative ctDNA. Although pathological risk factors are the routine marker to determine adjuvant agents and prognosis, the correlation is still disputable [28]. CEA and CA15-3 are used in post-operative monitoring for breast cancer, but the sensitivity and specificity are far from satisfying [5,6]. Of seven patients who experienced post-operative recurrence, elevated CEA/ CA15-3 was observed in three patients, and ahead of image findings of tumor recurrence in two patients, including one who was undetectable with ctDNA before imaging recurrence. As a comparison, ctDNA was detected at recurrence in all seven patients, and ahead of imaging recurrence in five patients. Through a combination of these markers, oncologists can better monitor the change of tumor burden and modify clinical management during follow-up.
A previous study focused on the TP53 mutations in blood ctDNA, and its association with efficacy and prognosis of neoadjuvant chemotherapy [13]. However, breast cancer is of high genomic heterogeneity. Even the top two mutant genes, TP53 and PIK3CA, do not cover an entire cohort [29][30][31][32]. As shown in this study, 37.5% (12/32) of patients had no TP53 or PIK3CA mutations in tumor tissue. Thus, a remarkable innovation of this study is the capture panel, which includes highly-frequently mutant genes across the main solid tumors. Using such approach, oncologists would no longer need to screen the applicable patients in advance, suggesting that the population who can benefit from ctDNA profiling may be enlarged. To eliminate the interference of background and hematopoietic mutations in the blood, we profiled the mutational landscape for each patient. Using two longitudinal tissue samples, temporal heterogeneity was revealed and could be avoided to some extent in the test for ctDNA. More longitudinal and multi-regional sampling can further eliminate the interference of spatial-temporal heterogeneity, but we think our strategy benefits patients most when considering physical invasion and economic cost.
Since neoadjuvant chemotherapy will exert a selective evolutionary pressure on the mutation spectrum or clonal structure of breast cancer patients [37][38][39][40], it is difficult to unify a threshold value to filter the mutations exactly. Therefore, we combined the mutations collectively for baseline and post-neoadjuvant samples as the reference to observe the consistency of the detection of peripheral blood and tissue mutations. One caveat was that it may confound the certain mutations better represented in baseline sample or post treatment sample. Additional limitations also persisted in this study, including its modest sample size and the lack of analysis about ctDNA detection in diverse patient subsets with different clinicopathological characteristics. Particularly, only four MBC were enrolled which restricted corresponding subgroup analysis. Besides, it was still confused whether post-operative ctDNA was derived from minimal residual disease or metastatic lesions, and simultaneous sampling of primary lesion, metastasis and peripheral blood might contribute to the problem. Despite aforementioned limitations, the results of this study offer a promising method to stratify patients before treatment by clinical management parameters, including surgical procedures and adjuvant agents, via noninvasive and highly sensitive blood ctDNA analysis. These results should be further explored and validated in larger-scale studies. We hope our efforts can help improve the clinical outcomes for patients undergoing neoadjuvant chemotherapy.

Clinical cohort
This prospective study recruited LABC patients at Peking Union Medical College Hospital (ClinicalTrials.gov Identifier: NCT02797652). All the patients were clinically evaluated by physical examination and imaging examination. Patients were enrolled when they met at least one of the following inclusion criteria: (1) tumors larger than 5 cm with regional lymphadenopathy (N1-N3); (2) cancers that involve the skin of the breast or the underlying muscles of the chest; (3) presence of regional lymphadenopathy (clinically fixed or matted axillary lymph nodes, or any infraclavicular, supraclavicular, or internal mammary lymphadenopathy) regardless of tumor stage; or (4) inflammatory breast cancer. The clinical stage was evaluated according to the American Joint Committee on Cancer TNM staging system for breast cancer [33]. In addition, some patients with metastatic breast cancer (MBC, stage IV) were also enrolled according to a proposal from at least three experienced clinicians. These patients had controllable and few detectable metastases, mainly in bone and internal mammary lymph nodes, which was in line with oligometastasis [41,42]. For preoperative therapy regimens, all the patients with HER2-cancer were given docetaxel/epirubicin/cyclophosphamide (TEC) regimen. Although dual HER2-blockade with both trastusumab and pertusumab was considered to be the optimal choice for HER2+ breast cancer in neoadjuvant setting, the unavailability of pertusumab in our country led to the adoption of docetaxel/carboplatin/ trastuzumab (TCbH) regimen for all the HER2+ patients. And all the HER2+ patients received one year trastusumab as adjuvant HER2-targeting therapy. The choice of surgery and medical agents was made by the treating clinician, who was blinded to the ctDNA profiling.
After surgery, the resected tissues were sent to the pathology laboratory at Peking Union Medical College Hospital. Both breast and lymph node tissue was estimated to determine ypT and ypN staging. The Miller-Payne classification system was used to evaluate the outcome of neoadjuvant chemotherapy [34]. We defined pathologic complete response (pCR) as simultaneous Miller-Payne grade 5 (no residual invasive cancer, but ductal carcinoma in situ may be present) and ypN0 (absence of cancer in lymph nodes). The concentrations of serum CEA and CA15-3 post-operatively or during follow-up were measured by the diagnostic laboratory at Peking Union Medical College Hospital. The cut-off value is 5 μg/L for elevated CEA and 30 U/mL for elevated CA15-3.
This study was registered at ClinicalTrials

Sample collection and DNA extraction
Before neoadjuvant treatment, tumor tissue biopsy was collected via core needle aspiration. In addition, we collected resected specimens during surgery. All tissue samples were fixed by formalin and then embedded in paraffin. At least 10-mL blood samples were collected in Streck tubes prior to chemotherapy, three cycles after the initiation of chemotherapy, 1-6 weeks after chemotherapy, 3-7 days after surgery, and at several follow-up time points. Less than 72 h after collection, peripheral blood was processed by centrifugation at 1600 × g for 10 min, then transferred to microcentrifuge tubes and centrifuged again at 16,000 × g for 10 min to remove cell debris. Plasma and peripheral blood lymphocytes (PBLs) were separated and stored at − 80 °C prior to DNA extraction.
Intact genomic DNA was extracted from tumor tissue and PBL using the QIAamp DNA mini kit and QIAamp DNA blood mini kit (Qiagen, Hilden, Germany), respectively. Circulating cell-free DNA (cfDNA) fragments were extracted from plasma using the QIAamp circulating nucleic acid kit (Qiagen, Hilden, Germany). Qubit fluorometer and Qubit dsDNA high sensitivity assay kit (Invitrogen, Carlsbad, CA, USA) were used to estimate the concentration of DNA. CfDNA fragment length was also estimated using an Agilent 2100 bioanalyzer and the DNA high sensitivity kit (Agilent Technologies, Santa Clara, CA, USA) to evaluate the contamination of large DNA fragments.

Sequencing library construction
Before library construction, genomic DNA was sheared into 200-250-bp fragments using the Covaris S2 instrument (Woburn, MA, USA). The DNA library preparation kit for Illumina (New England Biolabs, Ipswich, MA, USA) was used to prepare indexed next-generation sequencing libraries for genomic DNA fragments. As for cfDNA, after endrepairing and A-tailing reactions, well-designed adapters with unique identifiers were ligated to both ends of doublestranded cfDNA fragments. Then, PCR was used to generate sufficient fragments prior to hybridization. A previous article described additional detailed information about library construction [35].

Target region capture and next-generation sequencing
In hybrid capture procedure, we recruited custom-designed biotinylated oligonucleotide probes (IDT, Coralville, IA, USA) covering 1.09 Mbp of the human genome. Several processes were performed to embody the capture panel: (1) the most common driver genes across 12 common solid tumors [36] were selected and their entire exome regions were chosen; (2) genomic regions related with efficacy of chemotherapy, targeted drugs, and immunotherapy were added according to available clinical and pre-clinical studies; and (3) high-frequently mutant regions recorded in the Catalogue of Somatic Mutations in Cancer (COSMIC, http:// cancer. sanger. ac. uk/ cosmic) and The Cancer Genome Atlas (TCGA, https:// cance rgeno me. nih. gov/) were added. The overall included genes and genomic regions were shown in Table S1. We performed DNA sequencing using an Illumina 2 × 75-bp paired-end sequencing strategy on the HiSeq Sequencing System (Illumina, San Diego, CA, USA). The average sequencing depth was 800 × for tissue DNA, 200 × for PBL DNA, and 1500 × for cfDNA. Additional experimental information about target region capture and sequencing was also described in a previous article [35].

Raw data processing
Raw reads with adaptor sequences were first removed. Afterward, reads with more than 50% low-quality bases or more than 50% N bases together with their mate pair were further filtered. Residual reads were mapped to the reference human genome (hg19) using the Burrows-Wheel Aligner (http:// bio-bwa. sourc eforge. net/) with default parameters. Duplicate reads were identified and marked with Picard's Mark Duplicates tool (https:// softw are. broad insti tute. org/ gatk/ docum entat ion/ toold ocs/4. 0.3. 0/ picard_ sam_ markd uplic ates_ MarkD uplic ates. php). CfDNA data were clustered according to unique identifiers and the position of the template fragments. False-positive events introduced by PCR or sequencing procedures were removed according to clustered reads and mate pair reads. Local realignment and base quality recalibration were performed using the Gene Analysis Toolkit (https:// www. broad insti tute. org/ gatk/).
Insertions or deletions of small fragments (Indels) were called using MuTect2 with default parameters. Variants were removed if they were detected in matched control samples with three or more reads indicating Indels at the same location or in 40-bp flanking regions of experimental samples or residing near regions with low complexity or short tandem repeats.

Statistical analysis
Descriptive statistics were performed using SPSS 22.0 (IBM, Armonk, NY, USA). Differences in baseline characteristics between patients with positive and negative ctDNA at multiple time points were estimated using Fisher's exact test for categorical variables and Mann-Whitney (rank sum) test for continuous variables (SPSS 22.0, IBM, Armonk, NY, USA). Differences of cfDNA concentration and ctDNA allele frequency collected at different time points were calculated using One-Way ANOVA analysis. The primary endpoint was disease-free survival, measured from surgery to the first relapse or until the most recent follow-up of this study. A Kaplan-Meier analysis with the log-rank test was used to compare the disease-free survival (DFS) of different subsets of patients (GraphPad Prism 7, GraphPad Software, La Jolla, CA, USA). A p-value less than 0.05 indicated statistical significance.
Author contributions Totally 14 authors are listed in manuscript. YZ and YX are the main experimental designers and writers; CW, RY and LP have assisted the management of patients and collection of samples; YG, YZ, XZ and JB helped data arrangement and analysis; YG, XX, LY and XYoffered constructive suggestion for study design and analytical methods.

Data availability
The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.