BALB/c mice (male, 5–6 weeks old) were purchased from Orient Bio (Seongnam, Korea).

Four types of human non–small cell lung cancer (NSCLC) cell lines, H460, H1299 (Ras mutant), H1975, and HCC827 (epidermal growth factor receptor [EGFR] mutant), were obtained from ATCC (Manassas, VA). All cell lines were cultured with RPMI (Gibco, Carlsbad, CA) or Dulbecco’s modified Eagle’s medium (Gibco) containing 10% fetal bovine serum and 1% antibiotic-antimycotic (Thermo Fisher Scientific, Waltham, MA) and incubated at 37°C in 5% CO₂.

NSCLC cells, 1×10⁶ cells in a 100 μL suspension of Matrigel Basement Membrane Matrix (Corning, Tewksbury, MA), were subcutaneously injected into the hind legs of mice. The double-xenograft model used H1299, H1975, and HCC827 cells. Different mutant types of cancer cells were inoculated into the hind legs of each mouse. Tumor volumes were measured with calipers every 3 days and calculated according to the following formula: volume=D_Short²×D_Long/2. In both the single and two-tumor models, irradiation was delivered to the tumor-bearing legs of the mice in the irradiation groups using 6-MV photon beams from a linear accelerator (Varian Medical System, Palo Alto, CA) when the tumor volume reached 500 mm³ or 1,000 mm³. During irradiation, the mice were anesthetized by an intraperitoneal injection of tiletamine+zolazepam (50 mg/kg) and xylazine (10 mg/kg), under the approval of the Ministry of Food and Drug Safety. At certain times after irradiation, we sacrificed the mice and harvested their plasma and tumor tissues.

Immunohistochemistry (formalin-fixed paraffin-embedded sections) analysis of tumor tissues sections labeling F4/80 at 1/50 dilution (BM8, eBioscience, San Diego, CA). Heat mediated antigen retrieval citrate buffer. Anti-rat IgG, horseradish peroxidase conjugated (Dako, Glostrup, Denmark) was used as the secondary antibody. Hematoxylin was used as a counterstain.

Circulating DNA was extracted from plasma using a QIAamp Circulating Nucleic Acid Kit (Qiagen, Santa Clara, CA). Genomic DNA (gDNA) was isolated from blood samples using a QIAamp DNA Mini Kit (Qiagen). An AllPrep DNA/RNA Mini Kit (Qiagen) was used to purify gDNA and mRNA from tissues. DNA/RNA concentrations and purity were quantified using a Nanodrop 8000 UV-Vis spectrometer (Thermo Fisher Scientific) and a Picogreen fluorescence assay on a Qubit 2.0 fluorometer (Life Technologies, Waltham, MA). The fragment size distribution was measured using a 2200 TapeStation Instrument (Agilent Technologies, Santa Clara, CA).

To quantify isolated cfDNA, the long interspersed nuclear element-1 (LINE-1) locus was amplified by real-time polymerase chain reaction (PCR) using SYBR Green (Exiqon) according to the manufacturer’s protocols. The LINE-1 locus region was amplified by the following pair of PCR primers: hLINE1: 5′-TCA CTC AAA GCC GCT CAA CTA C-3′, 5′-TCT GCC TTC ATT TCG TTA TGT ACC-3′ and mLine1: 5′-GGA GGG ACA TTT CAT TCT CAT CA-3′, 5′-GCT GCT CTT GTA TTT GGA GCA TAG A-3′. The Ct (threshold cycle) values of the target genes were determined using LightCycler 480 software (Roche, Branchburg, NJ).

To validate the mutations in the NSCLC cell lines, droplet digital PCR (ddPCR) was performed using a QX200TM Droplet Digital PCR System (Bio-Rad, Hercules, CA) according to the manufacturer’s guidelines. The TaqMan ddPCR Liquid Biopsy Assays for NRAS p.Q61K and EGFR p.E746–A750del (Hs000000079_rm and Hs000000027_rm, Life Technologies) were used. The ddPCR data were analyzed using QuantaSoft software (Bio-Rad).

Purified gDNA was sonicated (7 minutes, 0.5% duty, intensity of 0.1, and 50 cycles/burst) into 150–200 bp fragments using a Covaris S2 (Covaris Inc., Woburn, MA). The tumor biopsy sample libraries were constructed using a SureSelect XT reagent kit, HSQ (Agilent Technologies) according to the manufacturer’s instructions. The peripheral blood lymphocytes and plasma DNA libraries were created using a KAPA Hyper Prep Kit (Kapa Biosystems, Woburn, MA). Briefly, after completing end repair and A-tailing according to the manufacturer’s protocol, we performed adaptor ligation at 4°C overnight using a pre-indexed PentAdapter (PentaBase ApS, Soendersoe, Denmark). Hybrid selection was performed using customized baits that targeted ~117 kb of the human genome, including exons from 38 cancer-specific genes (S1 Table). Sequencing libraries for whole exome sequencing were created using the SureSelect XT Human All Exon V5 kit (Agilent Technologies) and subsequently analyzed on a HiSeq 2500 system (Illumina, San Diego, CA) using the 100-bp paired-end mode of the TruSeq Rapid PE Cluster kit and TruSeq Rapid SBS kit (Illumina).

All liquid biopsy data were aligned to the hg19 reference using BWA-mem (v0.7.5) and analyzed as previously reported [9]. During processing, discordant pairs and off-target reads were filtered out. Whole transcriptome data for the xenograft mouse model were generated on an Illumina HiSeq 2500 system, and the sequences were aligned to a mixed genome of GRCh37 and GRChm38 by STAR 2.5.2b. Using RSEM 1.3.0, we quantified the aligned sequences into separate mouse and human transcriptomic expression levels and then merged the levels into each gene to calculate the transcripts per million values representing their gene expression. The cell-type deconvolution analysis was performed by MuSiC [10] for the RNA data from each mouse to characterize 5 distinct cell types in the bulk data. We measured the cell-type composition across each sample to compare the proportions of individual cell types.

First, all bases were subjected to Phred quality filtering using a threshold Q of 30, and only positions where total depths were above 500× were considered for variant identification. The error suppression method using unique molecular identifiers was carried out to select highly confident reads supporting a non-reference. Non-reference alleles present at a frequency greater than 1% in the matched germline DNA were removed. Otherwise, non-reference alleles were subjected to the binomial test to determine if a non-reference allele was significantly more abundant in plasma DNA than the matched germline DNA (Bonferroni adjusted p-value < 0.01). To minimize false-positives due to cross-contamination among multiplexed samples, we also excluded non-reference alleles if they were found as germline single nucleotide polymorphisms in other samples processed in a capture reaction or the same lane of a sequencing flow cell. Variant candidates with a high strand bias (90% if supporting reads ≥ 20; Fisher exact test, p-value < 0.1 if supporting reads < 20) were removed. Next, we performed a Z-test to identify variants that were present at a significantly higher frequency than the corresponding background errors in the normal samples (Bonferroni adjusted p-value < 0.05). We further applied the following threshold as previously reported [9]. Variants that passed the filter in Strelka2 were further considered.

All statistical data were analyzed using GraphPad Prism (GraphPad Software, La Jolla, CA). All data were evaluated using analysis of variance (ANOVA) with the Bonferroni post-hoc test or Wilcox test, and p < 0.05 was considered to indicate statistical significance.

To assess the clinical feasibility of TLB in lung cancer patients, we launched a prospective study for NSCLC patients undergoing definitive RT. The first protocol (IRB 2017-09-120) included patients with stage I–III NSCLC according to the 8th American Joint Committee on Cancer staging manual and pathologic confirmation who were inoperable because of their medical condition. They were recommended to receive definitive RT, and dose-fractionations were determined based on the tumor size, location, and stage. Approximately 10 mL of whole blood sample was obtained at simulation for RT planning. During RT, blood samples were drawn after 1, 2, and 3 days and 1 week, followed by weekly sampling. Posttreatment samples were drawn at 1 month, followed by 3-month intervals for 2 years with chest computed tomography. The second protocol (IRB 2018-05-155) included clinically diagnosed lung cancer patients who failed to have a pathologic diagnosis or for whom diagnostic procedures were significantly risky. All other procedures were the same as in the first protocol.

First, we examined if RT is an adequate modality to promote cfDNA release from a targeted tumor tissue using mouse xenograft model as schemed in Fig. 1. Due to the small amounts of plasma DNA obtained from mouse xenograft models, we took advantage of the LINE-1, a repetitive sequence in the human genome. Because ~100,000 of these elements exist in the human genome and a subset of these copies can be distinguished from mouse orthologues, a quantitative real-time PCR method measuring human LINE-1 (hLINE-1) was successfully implemented to quantify ctDNA (i.e., human DNA) in small volumes of plasma samples from mouse xenograft models in a previous study [11]. As an alternative way to quantify ctDNA, we performed a droplet digital PCR method to count DNA fragments carrying NRAS/EGFR mutations, because the hLINE-1 assay could be used to discriminate ctDNAs released from different clonality of tumors (or cancer cell lines) despite its high sensitivity for human DNA. To validate these methods to each other, we assessed correlation between hLINE-1 and NRAS/EGFR mutation copy numbers. The concentration of plasma hLINE-1 in the tumor-bearing mice correlated well with the NRAS Q61K mutation of the H1299 xenograft model and the EGFR E746-A750del mutation of the HCC827 xenograft model (Fig. 2A), demonstrating that both methods reliably quantified ctDNA in our mouse models.

To determine the relationship between tumor burden and ctDNA in the xenograft mouse models, we established mouse subcutaneous tumor models using various NSCLC cell lines. The xenograft models showed significant differences in tumor growth according to the various NSCLC cell lines (S2 Fig.), but the tumor growth rate and plasma hLINE-1 levels increased significantly in all cell lines when the tumor volume went above 900 mm³. Thus, the correlation between the tumor growth rate and the hLINE-1 level was consistent within each cell line and across the four cell lines used in this study (Fig. 2B).

ctDNA could be measured using the amount of hLINE-1 or variation in the injected cells at tumor volumes of more than 400 mm², confirming that ctDNA reflects the tumor burden.

The hLINE-1 concentration increased significantly after irradiation in all animal models except the H1975 tumor model (Fig. 3A). Because ctDNA might be increased when tumor actively grew and more tumor cells died, we plotted the hLINE-1 concentration against the tumor growth rate to negate the possibility that the elevated hLINE-1 level in the irradiated mice (compared to that in the controls) was simply caused by a difference in the tumor growth rate. As shown in S3 Fig., the data showed that irradiation increased the hLINE-1 independently of tumor growth rate in H460, H1299, and HCC827. Among the three radiation-responsive cell lines, H1299 and HCC827 were chosen to further elucidate the ctDNA secretion caused by irradiation because they displayed similar tumor growth rates and carry different mutations. To select an optimal dose of irradiation and a time point to evaluate the response to irradiation, H1299 and HCC827 xenograft mice were irradiated with three different doses (6, 12, and 18 Gy) and sacrificed at three different time points after irradiation (6, 18, and 24 hours). The level of ctDNA was elevated upon irradiation at all tested doses. The elevation of the ctDNA level was consistently maintained from 6 hours to 24 hours, as observed in both the H1299 and HCC827 xenograft models (Fig. 3B). On the other hand, after irradiation, the concentration of cfDNA showed a small increase or no change in both models. Therefore, in further experiments, the mice were irradiated with 12 Gy and sacrificed 18 hours after irradiation to analyze ctDNA.

In addition to hLINE-1 quantitative PCR, we performed a ddPCR assay to quantify DNA fragments carrying tumor-specific mutations and thereby double check the elevated ctDNA levels upon irradiation. In H1299 tumor-bearing mice, the NRAS mutation increased after irradiation, which was more prominent in mice with larger tumors. The EGFR mutation also increased after irradiation in HCC827 tumor-bearing mice (Fig. 3C). Overall, we found that irradiation increased the level of ctDNA in the mouse lung-cancer xenograft models.

Based on the increased level of ctDNA after irradiation, we hypothesized that radiation aimed at a particular tumor region would locally facilitate ctDNA release from the irradiated region and thus enable TLB. TLB was anticipated not only to improve sensitivity of ctDNA but also to overcome tumor heterogeneity. To emulate intratumor genetic heterogeneity, we established two-tumor model mice bearing H1299 and HCC827 tumors, one in each hind leg.

To investigate the hypothesis, we irradiated one tumor site in two-tumor model mice. When irradiation was delivered to the H1299 tumor, NRAS Q61K was expected to increase, whereas the EGFR E746-A750del present in the HCC827 tumor was not expected to be affected. Similarly, the EGFR mutation, but not the NRAS Q61K mutation, was expected to increase when the HCC827 tumor was irradiated (Fig. 1). In the two-tumor model with H1299 and HCC827 tumors, an increase in the target-specific mutation was observed: NRAS Q61K copies increased only when the H1299 tumor was irradiated, and EGFR E746-A750del copies increased only upon irradiation of the HCC827 tumor, suggesting preferential release of ctDNA from the irradiated tumors (Fig. 3D). These results confirm the feasibility of TLB in two-tumor mouse models.

To investigate the mechanism by which the tumor micro-environment modulates releases of cell-free DNA from cancer cells into blood vessels, we performed transcriptome analysis of xenograft tumor tissues comprising tumor cells and infiltrated normal cells. We obtained both tumor tissues in the two-tumor mouse model and generated next-generation sequencing-based RNA sequencing (RNA-seq) data.

As a result, a mean of 19,144,381 human and 2,806,448 mouse reads were uniquely mapped to a concatenated human and mouse genome, indicating that 13.3% of transcripts were originated from host cells (Fig. 4A). In human transcripts, differentially expressed genes with or without irradiation did not show a common pattern in each sample and were clustered with the characteristics of the injected cell lines. The pattern was relatively weak, but the similar results in mouse transcripts was indicating that the two tumor cell lines interacted differently with host cells (Fig. 4B). When we listed host genes whose expression were significantly correlated with the level of ctDNA, most of the adjusted genes were functionally related to the cell cycle, apoptosis, metabolism, and DNA repair. In the absence of irradiation, transcriptional signatures were clustered by the injected cell line, and when they were exposed to radiation, interestingly, we confirmed that they clustered with genes related to ctDNA release (Fig. 4C).

Because recent studies showed that radiation, particularly high-dose, hypo-fractionated administration [12], stimulates the immune system, we investigated changes in the composition of the immune cells. To estimate the proportions of cell types contained in the tumor microenvironment, we deconvoluted the bulk RNA-seq data (the sum of the gene expression profiles of relevant cell types weighted by the proportion of each cell type) using cell type–specific gene expression references [10]. After irradiation, the cell fraction of macrophages and monocytes among the immune cells increased slightly, and natural killer (NK) cells decreased (Fig. 4D), which correlated with the amount of ctDNA (Fig. 4E). Immunohistochemical analysis of mouse tumors using markers of the mouse macrophage population also confirmed that macrophages increased after irradiation (Fig. 4F).

Taken together, our data indicated that the secretion of ctDNA following RT was not only caused directly by cancer cells but also indirectly related to changes in the cell composition of the tumor microenvironment.

We launched a prospective study for patients undergoing definitive RT with or without a histologic diagnosis as schemed in Fig. 1. The first patient was a 78-year-old man with squamous cell carcinoma in the left upper lobe of his lung. The size of the primary tumor was 4.8 cm, and the clinical stage was T2bN1. He received definitive RT with a dose of 66 Gy in 22 fractions. Similar to our mouse tumor models, an elevation in the ctDNA level was observed 72 hours after the start of RT (Fig. 5A). The second patient was an 82-year-old man with a 1.9-cm tumor in the right lower lobe of his lung. Histologic diagnosis was judged to be difficult because of the tumor location, and he received hypo-fractionated RT with a dose of 64 Gy in 8 fractions. Despite the absence of a histologic diagnosis, several mutations were detected in the tumor DNA, and they increased after irradiation (Fig. 5B).

In total, we analyzed ctDNA in longitudinal blood samples from 11 patients (S4 Table). We observed a 2-fold increase in relative ctDNA levels 24–48 hours after the start of RT (Fig. 5C). Among the 11 patients, 10 (91%) showed temporary increase of the ctDNA within 72 hours after initiation of RT (Fig. 5D).