The Origin of Highly Elevated Cell-Free DNA in Healthy Individuals and Patients with Pancreatic, Colorectal, Lung, or Ovarian Cancer

In the peripheral blood of healthy individuals, the vast majority of cell-free DNA (cfDNA) is derived from cells of hematopoietic lineage, consistent with lymphoid and myeloid cell death as the predominant source of cfDNA (1–4). The first measurements of cfDNA from patients with cancer were performed more than 50 years ago (5–8). Although these studies were performed before the development of technologies that could characterize the DNA fragments in detail, it was clear that the concentration of DNA was often elevated in patients with cancer compared with healthy controls (5, 7, 9, 10). Numerous studies since then have confirmed that cfDNA in patients with cancer is often elevated and that patients with more advanced cancer are more likely to have higher cfDNA concentrations (9, 11). Sensitive studies of mutations and chromosome copy-number changes have shown that part of the cfDNA derived from patients with cancer is derived from neoplastic cells within the tumor (9–16). However, many recent studies of cfDNA are based on characteristics that are not entirely specific to neoplastic cells. These include assays based on cfDNA fragment ends, lengths, and enrichments in specific sequences such as promoters, as well as other epigenetic features that can be used to identify not only the presence of cancer but the cell type of origin. In general, previous studies have not been able to distinguish whether the great excess of cfDNA that is observed in some patients with cancer is derived from the neoplastic cells within the tumor rather than from normal epithelial cells surrounding the tumor that may have been damaged by the cancer (Supplementary Note 1 and refs. 3, 14–44). We sought to answer this question through a combination of genetic and epigenetic analysis of cfDNA, particularly in samples with very high concentrations of cfDNA (sample and analysis schema shown in Supplementary Fig. S1).

The normal concentration of cfDNA in healthy individuals generally ranges from 1 to 10 ng/mL of plasma, and the average concentration of cfDNA in the plasma of patients with cancer is higher than in healthy individuals (7, 9–11, 33–36). As an example, the distribution of cfDNA concentrations as measured by quantitative real-time PCR among 812 healthy individuals and 1,005 patients with cancer from a recently reported study (9) is shown in Fig. 1A. The mean concentration of cfDNA in the plasma of the normal individuals was 4.3 ± 8.6 ng/mL of plasma [2.9 ± 1.6 ng/mL, median ± median absolute deviation (MAD)], whereas the mean concentration of cfDNA in the plasma of patients with stage I–III cancer was 12.6 ± 18.1 ng/mL of plasma (6.30 ± 3.5 ng/mL, median ± MAD). The concentrations varied considerably with cancer type, with lung cancers having the lowest at 5.23 ± 6.4 ng/mL of plasma (3.3 ± 1.5 ng/mL, median ± MAD) and liver the highest at 46.0 ± 35.6 ng/mL of plasma (42.3 ± 29.4 ng/mL, median ± MAD; Fig. 1B). A particularly high concentration of cfDNA in liver cancers has been previously noted (3, 37, 38). There was a significant difference in cfDNA concentration between American Joint Commission on Cancer (AJCC) 7th edition stage I–III cancers as a whole (Supplementary Fig. S2A, P < 0.01), which varied within each cancer type (Supplementary Fig. S2B). Note that none of the 1,005 patients with cancer evaluated in Fig. 1B and Supplementary Fig. S2A and S2B had distant metastatic disease at the time plasma was taken, though it is well known that patients with metastatic disease have the highest cfDNA concentrations (7, 10, 11, 38).

The concentration of cfDNA in the plasma of the new cohort of normal individuals included in the present study was very similar to the cohort described in ref. 9 and Supplementary Note 2. For 64 normal individuals, the mean concentration of cfDNA was 6.0 ± 10.5 ng/mL of plasma (3.4 ± 0.8 ng/mL, median ± MAD; P = 0.15) was similar to that reported in ref. 9. For 178 patients with stage I–IV cancer evaluated in the new cohort studied here, the mean concentration of cfDNA was greater at 21.8 ± 26.5 ng/mL of plasma (P < 0.001), similar to that reported in ref. 9 at 11.8 ± 5.9, median ± MAD, reflecting the different cancer types and inclusion of patients with metastatic disease in the new cohort.

In 64 normal individuals, whole-genome bisulfite sequencing data showed that the vast majority of cfDNA arises from leukocytes regardless of the total concentration of cfDNA in the plasma (Supplementary Note 3; Fig. 2, blue dots; Supplementary Fig. S3; Supplementary Table S1). Neutrophils accounted for roughly two thirds of the leukocyte cfDNA, consistent with the ∼2:1 ratio of neutrophils to lymphocytes in the circulation of healthy individuals. Interestingly, the fraction of cfDNA contributed by B cells was not consistent with the fraction of B cells in the circulation of healthy individuals. B cells are expected to account for only 10% to 15% of blood lymphocytes, whereas T cells account for most of the remainder (80%; ref. 45). But B cell– and T cell–derived DNA accounted for 16.4% ± 12.0% and 17.5% ± 4.3%, respectively, of lymphocyte cfDNA in the circulation. This difference between the ratio of B to T cells expected in the circulation and the ratio of B cell–derived DNA to T cell–derived DNA in the circulation was highly significant (P < 0.001; Supplementary Note 4). Other minor tissue contributors to cfDNA were the liver, colon, heart, brain, and lung, accounting for 5.8% ± 6.6%, 4.0% ± 7.1%, 3.5% ± 3.4%, 3.1% ± 3.3%, and 2.4% ± 3.4% of the total, respectively (Fig. 2, blue dots; Supplementary Table S1). Deconvolution using nonnegative least squares linear regression (NNLS; refs. 40–42) instead of quadratic programming (QP; ref. 39) yielded nearly identical results (Supplementary Fig. S4).

Using a separate reference deconvolution matrix from Moss and colleagues (43) and QP (39) as the deconvolution algorithm, we found that leukocytes were again the predominant contributor to plasma cfDNA at 61.2% ± 23.4% (Fig. 3, blue dots; Supplementary Table S1), with neutrophils contributing the most cfDNA, followed by monocytes, natural killer (NK) cells, myeloid progenitors, B cells, and T cells. Similar to data obtained with the Sun and colleagues (3) deconvolution reference matrix, other minor contributors included colon epithelial cells at 7.2% ± 8.5% and hepatocytes at 6.7% ± 9.8%, respectively. Again, deconvolution using NNLS (40–42) instead of QP (39) provided very similar results (Supplementary Fig. S5). Contributions of overlapping cell types as determined by the Moss and colleagues (43) and Sun and colleagues (3) reference matrices were similar for total leukocytes, neutrophils, B cells, T cells, hepatocytes, colon epithelial cells, and brain (Supplementary Tables S1 and S2).

Deconvolution with the Loyfer and colleagues (44) matrix and the NNLS algorithm similarly showed that leukocytes were the predominant contributor to plasma cfDNA at 54.8% ± 20.3% (Supplementary Table S2), with blood granulocytes contributing the most cfDNA, followed by megakaryocytes, blood monocytes/macrophages, keratinocytes, erythrocyte progenitors, endothelial cells, NK cells, and T cells. Analysis of our healthy cohort showed tissue contributions very similar to those reported in the healthy patients evaluated by Loyfer and colleagues (44).

For individuals with nonelevated concentrations of cfDNA, the results described above are consistent with prior studies showing that most cfDNA comes from cells of lymphoid and myeloid lineage (2–4, 27). The novel aspect of the current study is the determination of these origins in individuals with elevated cfDNA. No prior studies had evaluated the tissue of origin of these elevated cfDNA concentrations in such individuals, and we hypothesized that such individuals might have had tissue-specific damage that accounted for their extremely high cfDNA levels. However, the results did not confirm our hypothesis; there was a linear correlation between the amount of DNA contributed by leukocytes and the total cfDNA concentration at all concentrations regardless of whether the Sun and colleagues (3), Moss and colleagues (43), or Loyfer and colleagues (44) reference matrices were used for deconvolution, or whether the plasma cfDNA methylation signature was deconvoluted by QP or NNLS (Fig. 2, blue dots, R² = 0.99, P < 0.001; Fig. 3, blue dots, R² = 0.96, P < 0.001; Supplementary Table S1). In other words, the great majority [average 79%, interquartile range (IQR) 74%–88%] of the cfDNA in healthy plasma, even when the total cfDNA concentration was more than 10× the normal level, arose from leukocytes. Similar fractions of DNA arising from neutrophils, B cells, and T cells were discovered regardless of the total cfDNA concentrations (R² = 0.99, 0.96, and 0.89, respectively, P < 0.001; Fig. 2, blue dots). The amount of liver and lung DNA was proportionately increased with total cfDNA concentration, in the same way as observed in healthy individuals without elevated cfDNA (Fig. 2, blue dots). Additionally, the same unexpectedly high contribution of B cells to total cfDNA was observed in normal individuals with high cfDNA concentrations as in those with low concentrations (Fig. 2, blue dots). In addition to total leukocytes, analysis using the Moss and colleagues (43) reference matrix highlighted the contribution of myeloid progenitor cells (R² = 0.91, P < 0.001), monocytes (R² = 0.86, P < 0.001), and neutrophils (R² = 0.79, P < 0.001) at all concentrations of total cfDNA (Fig. 3, blue dots).

We analyzed plasma from 178 patients with colorectal (N = 18), lung (N = 31), ovarian (N = 36), or pancreatic (N = 93) cancer to determine the source of their cfDNA. As in the normal individuals described above, leukocyte lysis during sample collection or processing or other contribution from high-molecular-weight DNA was excluded in all patients with cancer (Supplementary Fig. S3; Supplementary Table S1).

In patients with cancer, the tissue source of cfDNA (Fig. 2, red dots) was markedly similar to that in normal individuals (Fig. 2, blue dots). Using the reference matrix from Sun and colleagues (3) and deconvolution via QP, we discovered that 70.5% ± 13.7% (73.6% ± 5.4%, median ± MAD) of the cfDNA in these patients was contributed by leukocytes, with an average of 11.4% ± 11.4%, 5.9% ± 9.0%, 3.6% ± 2.8%, 3.1% ± 3.0%, 2.2% ± 3.7%, and 2.2% ± 2.7% contributed by liver, colon, brain, heart, lungs, and pancreas, respectively (Supplementary Table S1). Of the leukocyte DNA, about two thirds was derived from neutrophils in patients with cancer, just as in normal individuals (Supplementary Table S1). These results are consistent with previous studies on patients with cancer without elevated cfDNA concentrations (3). Deconvolution strategies using the Moss and colleagues (43) reference matrix or using NNLS instead of QP yielded similar results (Fig. 3, red dots, compared with Fig. 2, red dots; Supplementary Figs. S4 and S5; Supplementary Table S1). Analysis using the Loyfer and colleagues (44) approach also produced similar results, with total leukocytes contributing 56.6% ± 13.3% to the total cfDNA pool, with predominant contributions by blood granulocytes, megakaryocytes, blood monocytes/macrophages, and hepatocytes, just as in healthy individuals (Supplementary Table S2).

In patients with elevated concentrations of cfDNA, we expected that a major source of the large amounts of DNA in patients would be from the neoplastic cells and the surrounding nonneoplastic epithelial cells. This expectation was not confirmed by the experiment. As with normal individuals, there was a linear correlation between the amount of DNA contributed by leukocytes and the total cfDNA concentration (Fig. 2, red dots, R² = 0.92, P < 0.001; Fig. 3, red dots, R² = 0.82, P < 0.001).

Interestingly, 10 patients with cancer had approximately 5 ng/mL or more of their cfDNA derived from colonic epithelium (Supplementary Table S1). These concentrations were significantly different from those derived from normal individuals regardless of the total cfDNA concentration in the normal individuals (P < 0.001). Of the 10, 80% (8/10) of these patients had colorectal cancers. We sought to understand the origin of this epithelial DNA. In theory, it could have come from the neoplastic cells themselves or the surrounding nonneoplastic colorectal epithelial cells that had been destroyed by the cancer. It is well known that cancers destroy surrounding normal organ cells during the invasive process, and these dead or dying cells could in principle contribute to cfDNA (46–48). To distinguish between these two possibilities, tumor-specific mutations and copy-number alterations (CNA) in the cfDNA were used to determine the fraction of the cfDNA contributed by the neoplastic cells themselves.

We found a linear correlation between the fraction of cfDNA derived from colon epithelial cells and the fraction of cfDNA derived from neoplastic colon epithelial cells in patients with colorectal cancer (Supplementary Fig. S6, R² = 0.95, P < 0.001; Supplementary Note 5). The former was assessed by the whole-genome bisulfite sequencing of plasma cfDNA, whereas the latter was assessed by SafeSeqS analysis of mutations in the same plasma cfDNA samples, as described in Kinde and colleagues (49). The data in Supplementary Fig. S6 and Supplementary Table S1 show that the amount of cfDNA derived from all colonic epithelial cells (assessed by methylation) was similar to that expected from the contribution of the neoplastic colonic epithelial cells alone (as assessed by mutation; ref. 50). This conclusion was supported by copy-number analysis of the cfDNA (Supplementary Table S1).

Other conditions besides cancer have been associated with elevated cfDNA concentrations (51–53). For example, it has been shown that large increases in cfDNA occur 1 day after surgery (10, 54, 55). To investigate the source of the extra cfDNA in such patients, we obtained plasma samples approximately 24 hours after surgery from nine of the patients with pancreatic cancer included in Supplementary Table S1, all of whom had Whipple procedures for tumor resection. Prior to surgery, eight of the nine patients had total cfDNA concentrations in the normal range (Supplementary Fig. S7). Following surgery, there was a dramatic elevation of the total cfDNA, ranging from 2.3- to 18-fold (median of 8.7-fold) in these eight patients (Supplementary Fig. S7; Supplementary Table S1, P = 0.001). The only patient for whom the cfDNA concentration did not increase following surgery was the one (PANCA 1248) with elevated cfDNA (29.1 ng/mL of plasma) prior to surgery.

Bisulfite sequencing of plasma cfDNA in these eight patients revealed the following:

• (i) The amount of cfDNA from all evaluable tissue sources increased after surgery, though the additional contribution from the lungs, brain, esophagus, small intestines, pancreas, and heart appeared to be slightly elevated (P = 0.09, 0.13, 0.08, 0.09, 0.13, and 0.10, respectively; Fig. 4; Supplementary Table S1).

• (ii) The majority (average 57%; range, 42%–70%) of the total cfDNA after surgery was from leukocytes, and the predominant contributors to leukocyte cfDNA were neutrophils (average 73%; range, 67%–82% of the leukocyte cfDNA; Supplementary Fig. S8; Supplementary Table S1).

• (iii) For most tissues, the proportional representation of each evaluable tissue either decreased or changed only slightly after surgery—except for the liver (Supplementary Fig. S8; Supplementary Table S1). There was a striking increase (average 46-fold, range, 6- to 144-fold) in the fraction of total cfDNA derived from hepatocytes following surgery, in marked contrast to the other tissues (Supplementary Fig. S8). The amount of the “neo-cfDNA” can be defined and calculated by subtracting the amount of cfDNA present before surgery from the amount of cfDNA present after surgery. This calculation showed that hepatocytes contributed an average of 38.2% (range, 23.9%–57.6%) and leukocytes contributed an average of 48.4% (range, 23.3%–67.1%) of the neo-cfDNA.

Based on the clinical history of PANC 696 described above, as well as on previous data (3, 52), we thought it likely that much of the extra cfDNA following surgery was due to liver damage. We were able to obtain standard measurements of liver function using alanine transaminase (ALT) and aspartate aminotransferase (AST) levels in five of the eight patients whose cfDNA concentrations increased following surgery. AST and ALT levels substantially increased in all five patients (P < 0.05; Supplementary Fig. S9). Notably, in the one patient (PANCA 1248) whose total cfDNA did not increase after surgery, AST levels were already increased prior to surgery, unlike the other patients assessed (Supplementary Table S1).

The results of this study lend support to previous observations about the origins of cfDNA in healthy individuals and patients with cancer who have normal to slightly elevated concentrations of cfDNA. Additionally, the results of this study lead to several important conclusions about the origin of excess cfDNA in patients with greatly elevated cfDNA concentrations:

• (i) In patients with colorectal, lung, ovarian, and pancreatic cancers with high concentrations of cfDNA in the present study, the increased cfDNA does not primarily come from either the neoplastic cells within the cancer or the adjacent nonneoplastic epithelial cells.

• (ii) Instead, the increased cfDNA in these patients with cancer comes largely from leukocytes, primarily neutrophils.

• (iii) The elevated cfDNA in patients with cancer studied is attributable to a systemic effect. It is not just neutrophils, but also B and T lymphocytes, and in some cases hepatocytes, colon epithelial cells, and lung epithelial cells, that release more DNA into the circulation when cfDNA concentrations are elevated.

• (iv) Similarly, the elevated cfDNA that routinely occurs following surgery of patients with pancreatic cancer arises from a systemic effect, resulting in the release of cfDNA from leukocytes but in this case also from hepatocytes.

• (v) The cfDNA contributed by leukocytes is often associated with an overrepresentation of B cells compared with T cells regardless of disease state and cfDNA concentration.

One of the major questions raised by our data is the nature of the systemic factor(s) that are responsible for increasing the contributions of cfDNA from all major tissue sources of cfDNA (56–73). One possibility is that the systemic factor(s) are one or more of the myriad of proteins and other molecules known to be secreted by neoplastic cells (74, 75), or released upon the death of cancer cells in situ (46, 75). Another possibility is that these factors come from endothelial cells within the cancers. There is convincing evidence indicating that the tumor vasculature is abnormal (76, 77) and endothelial cells are in direct contact with the systemic circulation. Inflammatory cells within the tumor could also release cytotoxic products (78). A completely different but enticing possibility is that cell turnover is normal in these patients, but clearance of cfDNA is abnormal. We hope that the results of this study will stimulate research to identify the biochemical basis for the pronounced elevation of cfDNA observed in patients with cancer and in other clinical scenarios.

All individuals participating in the study provided written informed consent after approval by the institutional review board (IRB) at the patient's participating institution (including Johns Hopkins IRB00075499 and Melbourne Health Human Research Ethics Committee 2011.225), and the study complied with the Health Insurance Portability and Accountability Act and the Declaration of Helsinki. Peripheral blood was collected in K2-EDTA tubes after informed consent was obtained and prior to and/or 24 hours after patients underwent surgical resection. General demographics, surgical pathology, and AJCC stage (7th) were documented. The cohort is outlined in Supplementary Fig. S1. The healthy cohort consisted of peripheral blood samples obtained from 64 individuals of median age 48.5 years (IQR, 28–58 years) with no history of cancer. The cancer and healthy control samples were processed in an identical manner. Plasma samples from 18 patients with colorectal cancer, 31 patients with lung cancer, 36 patients with ovarian cancer, and 93 patients with pancreatic cancer were included in the study (median age 67 years; IQR, 56–74 years).

The 242 individuals included in this study were chosen from cfDNA samples collected for studies described in ref. 48 and similar ongoing studies to evaluate the use of cfDNA for the earlier detection of cancer in patients prior to surgery or any other form of therapy. All individuals for whom sufficient plasma was available for construction of libraries for whole-genome sequencing of bisulfite-treated DNA were considered. Any individual with a cfDNA concentration >15 ng/mL of plasma from this collection was chosen for analysis. Additionally, there were two different blood samples available from nine of the 21 patients with pancreatic cancer, one collected prior to surgery and the other collected approximately 24 hours after surgery, and these were chosen for analysis. Finally, normal individuals with cfDNA concentrations <15 ng/mL of plasma, as well as patients with cancer with cfDNA concentrations <15 ng/mL, were chosen randomly. DNA from each of these 242 patients (251 total plasma samples) was purified with a BioChain cfDNA Extraction Kit (BioChain, cat. #K5011610) using the manufacturer's recommended protocol. DNA from peripheral white blood cells was purified with the QIAsymphony DSP DNA Midi Kit (Qiagen, cat. #937255) as specified by the manufacturer. cfDNA was quantified using qPCR using Sso Advanced SYBR Green Supermix (Bio-Rad, cat. # 1725271) as directed by the manufacturer and using the following primers:

• 5″-CACACAGGAAACAGCTATGACCATGGGTAACAGCTTTATCTATTGACATTATGC-3″

• 5″-CGACGTAAAACGACGGCCAGTNNNNNNNNNNNNNNAAACTTCATGCTTCATCTAGTCAGC-3″

National Institute of Standards and Technology (NIST) human DNA quantification standard NIST SRM 2372a, diluted to 1 ng/mL, served as the reference standard. cfDNA or NIST 2372a DNA (2.5 μL) was added to 97.5 μL of 1:1,000 SYBR Green I diluted in 1× PBS. Amplification and fluorescence detection conditions were as follows: one cycle of 98°C for 120 seconds and then 30 cycles of 98°C for 10 seconds, 57°C for 120 seconds, and 72°C for 120 seconds.

For libraries prepared using the Accel-NGS Methyl-Seq DNA Library Kit (Swift BioSciences, cat. #30024; Supplementary Table S1), the EZ DNA Methylation Kit (Zymo Research, cat. #D5001) was used to prepare DNA samples as follows. DNA was denatured in dilute M-Dilution buffer at 37°C for 15 minutes and then bisulfite-converted in the dark at 50°C for 16 hours before being placed on ice for 10 minutes (79). After a single wash with M-Wash buffer, the sample was desulphonated for 15 minutes at room temperature. The sample was washed twice in M-Wash buffer and then eluted in the Zymo Elution Buffer and stored at –20°C (49). Sequencing libraries were then prepared using the Accel-NGS Methyl-Seq DNA Library Kit, with nine PCR cycles used at the indexing stage. Samples assessed using MethylSaferSeqS (Supplementary Table S1) were prepared as described in detail by Wang and colleagues (80), in which library preparation is performed prior to bisulfite treatment (50). Tissue-specific methylation status as assessed by the two library preparation methods produced indistinguishable results (50). Each library was paired-end sequenced to 150 bp on a single lane of an Illumina HiSeq 4000 instrument. Reads passing Illumina CASAVA Chastity filters were used for subsequent analysis.

Illumina adapters and bases with quality scores below 25 were trimmed from the head and tail of each read using Trimmomatic (81). To improve mapping efficiency by reducing spurious mutations introduced by end repair, 15 bp were additionally cropped from the tail of Read 1 and the head of Read 2 using Trimmomatic per Swift's recommendations. BSMAP was used to align each paired-end read to the bisulfite-converted hg19 genome, and the average methylation at each CpG was computed using BSMAP's methratio.py script (82).

The average contribution of 12 tissue types (liver, lungs, colon, small intestines, pancreas, adrenal glands, esophagus, heart, brain, T cells, B cells, and neutrophils) to the total cfDNA pool was determined using 5,653 differentially methylated 500-bp regions described by Sun and colleagues (3). In brief, the approach was bioinformatically based on whole-genome bisulfite sequencing of normal DNA from the liver, lungs, esophagus, heart, pancreas, colon, small intestines, adrenal glands, brain, and T cells, which was retrieved from the Human Epigenome Atlas from the Baylor College of Medicine. The bisulfite sequencing data for B cells and neutrophils are from Hodges and colleagues (83). All CpG islands (CGI) and CpG shores on autosomes were assessed for potential inclusion into the methylation marker set. CGIs and CpG shores on sex chromosomes were not used, so as to minimize potential variations in methylation levels related to the sex-associated chromosome dosage difference in the source data. CGIs were downloaded from the University of California Santa Cruz (UCSC) database (genome.ucsc.edu; 27,048 CGIs for the human genome; ref. 84), and CpG shores were defined as 2-kb flanking windows of the CGIs (85). Then, the CGIs and CpG shores were subdivided into non­overlapping 500-bp units, and each unit was considered a potential methylation marker.

The methylation densities (i.e., the percentage of CpGs being methylated within a 500-bp unit) of all the potential marker loci were compared between the 12 tissue types. Using the methylation profiles of the 12 tissue types, two types of methylation markers were identified. Type I markers refer to any genomic loci with methylation densities that are 3 SDs below or above in one tissue compared with the mean level of the 12 tissue types. Type II markers are genomic loci that demonstrate highly variable methylation densities across the 12 tissue types. A locus is considered highly variable when (i) the methylation density of the most hypermethylated tissue is at least 20% higher than that of the most hypomethylated one, and (ii) the SD of the methylation densities across the 12 tissue types when divided by the mean methylation density (i.e., the coefficient of variation) of the group is at least 0.25. To reduce the number of potentially redundant markers, only one marker would be selected in one contiguous block of two CpG shores flanking one CGI. The genomic locations of the type I and II markers used in this study can be found in Supplementary Table S1 in Sun and colleagues (3).

As described in the work by Sun and colleagues (3), the mathematical relationship between the methylation densities of the different methylation markers in plasma and the corresponding methylation markers in different tissues can be expressed as

where |${\overline {MD} }_{\rm{i}}$| represents the methylation density of the methylation biomarker i in the plasma; p_k represents the proportional contribution of tissue k to the plasma; and MD_ik represents the methylation density of the methylation biomarker i in tissue k. The aim of the deconvolution process was to determine the proportional contribution of tissue k to the plasma, namely, p_k, for each member of the panel of tissues. QP (39) was used to solve the simultaneous equations. A matrix was compiled including the panel of tissues and their corresponding methylation densities for each methylation marker on the combined list of type I and II markers (a total of 5,653 markers). The program input a range of p_k values for each tissue type and determined the expected plasma DNA methylation density for each marker. The tested range of p_k values should fulfill the expectation that the total contribution of all candidate tissues, namely, the liver, neutrophils, and lymphocytes for this study, to plasma DNA would be 100% and the values of all p_k would be nonnegative. The program then identified the set of p_k values that resulted in expected methylation densities across the markers that most closely resembled the data obtained from the plasma DNA bisulfite sequencing.

The total contribution from T cells and B cells was regarded as the contribution from the lymphocytes, and the total contribution from leukocytes was regarded as the contribution from the lymphocytes and neutrophils. To obtain absolute levels of cfDNA (ng/mL) per cell type, the resulting contribution was multiplied by the total concentration of cfDNA present in the sample.

The average contribution of 24 tissue types (neutrophils, monocytes, CD4 T cells, CD8 T cells, B cells, NK cells, myeloid progenitors, adipocytes, cortical neurons, hepatocytes, lung cells, pancreatic acinar cells, pancreatic duct cells, vascular endothelial cells, colon epithelial cells, left atrium, bladder, breast, head and neck/larynx, kidney, prostate, thyroid, upper gastrointestinal, uterus/cervix) to the total cfDNA pool was determined using 7,890 differentially methylated CpGs, as described in Moss and colleagues (43). In brief, all DNA methylation profiles were determined either on the Illumina Infinium Human Methylation 450K or using the EPIC BeadChip arrays. DNA methylation data for white blood cells (neutrophils, monocytes, B cells, CD4⁺ T cells, CD8⁺ T cells, and NK cells, n = 6 each) were downloaded from GSE110555 (EPIC; ref. 86). Data for myeloid progenitors (n = 5) were downloaded from GSE63409 (450K; ref. 87), and data for left atrium (n = 4) were downloaded from GSE62727 (450K; ref. 88). Data for bladder (n = 19), breast (n = 98), cervix (n = 3), colon (n = 38), esophagus (n = 16), oral cavity (n = 34), kidney (n = 160), prostate (n = 50), rectum (n = 7), stomach (n = 2), thyroid (n = 56), and uterus (n = 34) were downloaded from The Cancer Genome Atlas (89). DNA methylation data for adipocytes (n = 3, 450K), hepatocytes (n = 3, 450K and EPIC), alveolar lung cells (n = 3, EPIC), neurons (n = 3, 450K and EPIC), vascular endothelial cells (n = 2, EPIC), pancreatic acinar cells (n = 3, 450K and EPIC), pancreatic duct cells (n = 3, 450K and EPIC), and colon epithelial cells (n = 3, EPIC) were generated by Moss and colleagues (43) and can be requested from the authors.

To analyze DNA methylation samples composed of admixed methylomes from various cell types, the authors approximated the plasma cfDNA methylation profile as a linear combination of the methylation profiles of cell types in the reference atlas. According to this model, the relative contributions of different cell types to plasma cfDNA can be determined using NNLS as described in refs. 40–42. To select candidate CpGs, the authors of refs. 40–42 first excluded CpGs whose variance across the entire methylation atlas was below 0.1% or was missing. They then selected the K = 100 most specific hypermethylated CpGs for each cell type, denoting the methylation matrix X, composed of N rows (CpGs) by d columns (cell types). They then divided each row (the methylation pattern of one CpG over all cell types) by its sum:

For each cell type j, they identified the top K hypermethylated CpGs with the highest X_i,j values. To identify uniquely hypomethylated CpGs, they performed a similar process for the reversed methylation matrix (1−X). Finally, for each cell type, they included both the top K hypermethylated and the top K unmethylated CpGs in the reference matrix. To this set of CpGs, they added neighboring CpGs up to 50 bp. Pairwise-specific CpGs were iteratively selected as follows: Given the current set S of CpGs, they projected the reference atlas on those coordinates and calculated the Euclidean distances between pairs of cell types. Once the closest pair of cell types was identified, they selected the CpG site where they differed the most and added it into the set S. This process was iteratively repeated, focusing on the most confusing pair of cell types in each iteration. Admixing experiments, similar to those performed in Sun and colleagues (3), were performed using buffy coat bisulfite sequencing data mixed with liver, lung, colon epithelial cell, or left atrium bisulfite sequencing data, showing excellent agreement between predicted fraction and actual fraction (as shown in Supplementary Fig. S10A–S10D).

A custom Python script adapted from the nnls package in MATLAB and described in Moss and colleagues (43) was used to perform NNLS (40–42) to calculate the relative contribution of each cell type to a given sample. Given a matrix X of reference methylation values with N CpGs and d cell types, and a vector Y of methylation values of length N, nonnegative coefficients β were identified by solving argmin _β∥Xβ − Y∥₂, subject to β ≥ 0. The resulting β was adjusted to have a sum of 1, where for each β_j was defined as:

To obtain absolute levels of cfDNA (ng/mL) per cell type, the resulting βj′ was multiplied by the total concentration of cfDNA present in the sample, as measured by quantitative PCR.

A similar analysis using NNLS with an expanded matrix of 39 cell types is described in Loyfer and colleagues (44).

RealSeqS was used to test the plasma samples for evidence of aneuploidy and contamination with high-molecular-weight DNA derived from leukocytes that were lysed during venipuncture or blood processing (12). RealSeqS uses a single primer pair to amplify ∼750,000 loci scattered throughout the genome (12). PCR was performed in 25-μL reactions containing 7.25 μL of water, 0.125 μL of each primer, 12.5 μL of NEBNext Ultra II Q5 Master Mix (New England Biolabs, cat. #M0544S), and 5 μL of DNA. The cycling conditions were one cycle of 98°C for 120 seconds and then 15 cycles of 98°C for 10 seconds, 57°C for 120 seconds, and 72°C for 120 seconds. Each plasma DNA sample was assessed in eight independent reactions, and the amount of DNA per reaction varied from ∼0.1 to 0.25 ng. A second round of PCR was then performed to add dual indexes (barcodes) to each PCR product prior to sequencing, as described in Douville and colleagues (12). The second round of PCR was performed in 25-μL reactions containing 7.25 μL of water, 0.125 μL of each primer, 12.5 μL of NEBNext Ultra II Q5 Master Mix (New England Biolabs, cat. #M0544S), and 5 μL of DNA containing 5% of the PCR product from the first round. The cycling conditions were one cycle of 98°C for 120 seconds and then 15 cycles of 98°C for 10 seconds, 65°C for 15 seconds, and 72°C for 120 seconds. Amplification products from the second round were purified with AMPure XP beads (Beckman, cat. #a63880), as per the manufacturer's instructions, prior to sequencing. As noted above, each sample was amplified in eight independent PCRs in the first round. Each of the eight independent PCRs was then reamplified using index primers in the second PCR round. The sequencing reads from the 8 replicates were summed for the bioinformatic analysis but could also be assessed individually for quality control purposes. Massively parallel sequencing was performed on an Illumina HiSeq 4000. During the first round of PCR, degenerate bases at the 5′ end of one of the primers were used as molecular barcodes [unique identifiers (UID)] to uniquely label each DNA template molecule (13). This ensured that each DNA template molecule was counted only once, as described in Kinde and colleagues (39). In all instances for RealSeqS in this article, the term “reads” refers to uniquely identified reads (UIDs). If multiple reads had the same UID, at least 50% of the reads were required to map to the same genomic location. Reads with the same UID but with discordant genomic locations were discarded from analysis.

After massively parallel sequencing, gains or losses of each of the 39 chromosome arms covered by the assay were determined using a bespoke statistical learning method (13). A support vector machine (SVM) was used to discriminate between aneuploid and euploid samples. The SVM was trained using 2,651 aneuploid samples and 1,348 euploid plasma samples to yield a “genome-wide aneuploidy score.” Samples were scored as positive when the genome-wide aneuploidy score was >0.441.

Plasma samples were also analyzed for genomic DNA contamination using RealSeqS. RealSeqS enables the detection of genomic DNA by virtue of the differently sized amplicons generated during PCR amplification (12). Because the average size of cfDNA is ∼160 to 180 bp, almost all the ∼750,000 amplicons are present in an average cfDNA sample. However, there were 1,241 amplicons of size 200 to 500 bp, which represent contamination by genomic DNA. Coverage at these long amplicons is proportional to the background rate of genomic DNA contamination, as described in ref. 13. In samples containing >15 ng of DNA per mL of plasma in which RealSeqS data were not available, an Agilent BioAnalyzer System was used to evaluate the fraction of DNA >500 bp.

For patients with colorectal cancer, a panel of 15 genes was designed to find mutations in DNA from primary tumors, as described in Tie and colleagues (90). This panel enabled the detection of at least one mutation in 98% of colorectal cancer samples tested (90). The mutation with the highest mutant allele frequency in the primary tumor was then used to assess plasma DNA, as described by Tie and colleagues (90). The SafeSeqS approach, using unique identifiers (UIDs, aka molecular barcodes), was then used to assess the plasma DNA for the mutation of interest (12). For patients with pancreatic cancer, plasma was directly assessed with SaferSeqS primers (9, 91) for mutations at codons 12, 13, 59, 60, and 61, as >95% of pancreatic cancers harbor a mutation at one of these positions (92).

ichorCNA version 3.2 was downloaded on August 25, 2022, and applied to whole-genome sequencing data on bisulfite-treated cfDNA. Tumor fraction estimates were based on CNA of 500-kb intervals using default parameters. The lower limit of detection was considered to be 3% based on data from ref. 93.

A 𝛘² test was used to compare the number of individuals with elevated concentrations of cfDNA in eight cancer types compared with healthy persons. A one-way ANOVA was used to compare the number of individuals with elevated concentrations of cfDNA in eight cancer types by AJCC 7th edition stage. The Pearson correlation coefficient was used to determine the relationship between total cfDNA concentration and relative contribution from individual tissues, and a t statistic was used to determine statistical significance. A Student two-tailed t test was used to compare the total concentration of cfDNA before and after surgery and AST and ALT levels before and after surgery. A P value ≤0.05 was considered statistically significant. In general, the numbers reported in this article are accurate to two significant digits. Three digits are reported in some instances solely because three digits were convenient for comparing various drafts of the article during its writing and editing.

Data on methylation (bisulfite sequencing) and CNAs in plasma DNA are deposited in the European Genome-phenome Archive (EGAS00001005400). Similarly, data on mutations in plasma are available from the European Genome-phenome Archive (EGAS00001002764 and EGAS00001002444). Commercial use remains restricted due to Johns Hopkins Medicine's legal requirements.

C. Douville reports personal fees from Exact Sciences during the conduct of the study; personal fees from Belay Diagnostics outside the submitted work; and a patent for RealSeqS licensed and with royalties paid from Exact Sciences. Y. Wang reports a patent for MethylSaferSeqS pending to Johns Hopkins, and is a consul­tant for Thrive Earlier Detection/Exact Sciences. A.H. Pearlman reports grants from the Lustgarten Foundation for Pancreatic Cancer Research, the Virginia and D.K. Ludwig Fund for Cancer Research, The Sol Goldman Center for Pancreatic Cancer Research, The Marcus Foundation, the John Templeton Foundation, and the NIH during the conduct of the study; grants from the Lustgarten Foundation for Pancreatic Cancer Research, the Virginia and D.K. Ludwig Fund for Cancer Research, The Sol Goldman Center for Pancreatic Cancer Research, The Marcus Foundation, the John Templeton Foundation, and the NIH outside the submitted work; and a patent for 15/560,241 pending, a patent for 16/614,005 pending, a patent for 63/059,638 pending, a patent for 17/783,506 pending, a patent for PCT/US2022/019980 pending, a patent for PCT/US2021/061453 pending, a patent for 63/119,753 pending, and a patent for PCT/US23/12290 pending. J.D. Cohen reports other support from Haystack Oncology and personal fees from Exact Sciences outside the submitted work. J. Tie reports personal fees from Haystack Oncology, AstraZeneca, Pierre Fabre, MSD, Bristol Myers Squibb, BeiGene, Seres Therapeutics, and Takeda outside the submitted work. C. Bettegowda reports personal fees from Haystack Oncology, Privo Technologies, Depuy-Synthes, Galectin Therapeutics, and Bionaut Labs, and other support from OrisDx and Belay Diagnostics outside the submitted work. R.H. Hruban reports personal fees from Thrive Earlier Detection during the conduct of the study. C. Tomasetti reports grants from the John Templeton Foundation during the conduct of the study; other support from Exact Sciences, Haystack Oncology, and PrognomiQ outside the submitted work; a patent for CancerSEEK with royalties paid from Exact Sciences, a patent for RealSeqS with royalties paid from Exact Sciences, a patent for BestSeqS pending, and a patent for BarcSeqS pending; and consulting for Bayer. P. Jiang reports grants from the Innovation and Technology Commission of the Hong Kong SAR Government (InnoHK Initiative) during the conduct of the study; personal fees from KingMed Future, Take2 Health, and Illumina and nonfinancial support from DRA outside the submitted work; and many patents/patent applications related to cfDNA molecules pending, issued, licensed, and with royalties paid from Grail, Illumina, DRA, Take2 Health, and Xcelom. K.C.A. Chan is a director of Take2, DRA, Insighta, and Prenetics; holds equities in Take2, DRA, Insighta, Prenetics, and Illumina; and holds patents on molecular diagnostics, and parts of the patent portfolio are licensed to Take2, DRA, Insighta, Grail, Illumina, and Labcorp and he receives royalties from these companies. Y.M.D. Lo is supported by the Innovation and Technology Fund under the InnoHK Initiative and by an endowed chair from the Li Ka Shing Foundation; holds equities in Insighta, DRA Limited, and Take2; has patents or patent applications that have been licensed to and receives royalties from Grail, Illumina, Labcorp, Xcelom, DRA Limited, Take2, and Insighta; and is a scientific cofounder of Grail. N. Papadopoulos reports grants from the NIH and the Commonwealth Foundation during the conduct of the study; personal fees and other support from Thrive Earlier Detection, an Exact Sciences Company, other support from Personal Genome Diagnostics, a Labcorp Company, Haystack Oncology, CAGE Pharma, NeoPhore, and ManaT Bio, and personal fees from Vidium outside the submitted work; a patent for MethylSaferSeqS: a method for simultaneous detection of genetic and epigenetic changes pending and a patent for bisulfite-converted duplexes for the strand-specific detection and quantification of rare mutations pending; and some of the companies named above, as well as other companies, have licensed previously described technologies related to the work described in this article from Johns Hopkins University, and N. Papadopoulos is an inventor on some of these technologies. Licenses to these technologies are or will be associated with equity or royalty payments to the inventors as well as to Johns Hopkins University. Patent applications on the work described in this article may be filed by Johns Hopkins University. The terms of all these arrangements are being managed by Johns Hopkins University in accordance with its conflict-of-interest policies. K.W. Kinzler reports grants from the Lustgarten Foundation for Pancreatic Cancer Research, the Virginia and D.K. Ludwig Fund for Cancer Research, The Sol Goldman Center for Pancreatic Cancer Research, The Marcus Foundation, the John Templeton Foundation, and the NIH during the conduct of the study; personal fees and other support from Thrive Earlier Detection, an Exact Sciences Company, and other support from Personal Genome Diagnostics, CAGE Pharma, NeoPhore, ManaT Bio, and Haystack Oncology outside the submitted work; a patent for MethylSaferSeqS: a method for simultaneous detection of genetic and epigenetic changes pending and a patent for bisulfite-converted duplexes for the strand-specific detection and quantification of rare mutations pending; and some of the companies named above, as well as other companies, have licensed previously described technologies related to the work described in this article from Johns Hopkins University, and K.W. Kinzler is an inventor on some of these technologies. Licenses to these technologies are or will be associated with equity or royalty payments to the inventors as well as to Johns Hopkins University. Patent applications on the work described in this article may be filed by Johns Hopkins University. The terms of all these arrangements are being managed by Johns Hopkins University in accordance with its conflict-of-interest policies. B. Vogelstein reports grants from the Lustgarten Foundation for Pancreatic Research, the Virginia and D.K. Ludwig Fund for Cancer Research, The Sol Goldman Center for Pancreatic Cancer Research, The Marcus Foundation, the John Templeton Foundation, and the NIH during the conduct of the study; other support from Thrive Earlier Detection, an Exact Sciences Company, Personal Genome Diagnostics, CAGE Pharma, NeoPhore, ManaT Bio, Haystack Oncology, and Catalio Capital Management outside the submitted work; a patent for MethylSaferSeqS: a method for simultaneous detection of genetic and epigenetic changes pending and a patent for bisulfite-converted duplexes for the strand-specific detection and quantification of rare mutations pending; and some of the companies named above, as well as other companies, have licensed previously described technologies related to the work described in this article from Johns Hopkins University, and B. Vogelstein is an inventor on some of these technologies. Licenses to these technologies are or will be associated with equity or royalty payments to the inventors as well as to Johns Hopkins University. Patent applications on the work described in this article may be filed by Johns Hopkins University. The terms of all these arrangements are being managed by Johns Hopkins University in accordance with its conflict-of-interest policies. No disclosures were reported by the other authors.

A.K. Mattox: Conceptualization, resources, data curation, software, formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. C. Douville: Formal analysis, writing–review and editing. Y. Wang: Data curation, writing–review and editing. M. Popoli: Data curation, writing–review and editing. J. Ptak: Resources, writing–review and editing. N. Silliman: Resources, writing–review and editing. L. Dobbyn: Resources, writing–review and editing. J. Schaefer: Resources, writing–review and editing. S. Lu: Data curation, writing–review and editing. A.H. Pearlman: Data curation, writing–review and editing. J.D. Cohen: Writing–review and editing. J. Tie: Resources, writing–review and editing. P. Gibbs: Resources, writing–review and editing. K. Lahouel: Software, writing–review and editing. C. Bettegowda: Supervision, writing–review and editing. R.H. Hruban: Resources, writing–review and editing. C. Tomasetti: Formal analysis, writing–review and editing. P. Jiang: Formal analysis, writing–review and editing. K.C.A. Chan: Formal analysis, writing–review and editing. Y.M.D. Lo: Formal analysis, writing–review and editing. N. Papadopoulos: Formal analysis, writing–review and editing. K.W. Kinzler: Formal analysis, writing–review and editing. B. Vogelstein: Conceptualization, resources, data curation, software, formal analysis, supervision, funding acquisition, validation, investigation, visualization, methodology, writing–original draft, project administration, writing–review and editing.

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.