Direct analysis of transcription factor protected cfDNA in plasma by ChIP-seq: measurement of altered CTCF binding in cancer is a novel biomarker for liquid biopsy

Abstract

Background

The aggregate presence of cell free CTCF-DNA (cfCTCF-DNA) nucleoproteins in plasma has been reported. But the occupancy of individual CTCF binding sites in plasma cfCTCF-DNA has not been studied.

Results

For the first time, we have isolated endogenous plasma cfCTCF-DNA by chromatin immunoprecipitation (ChIP) and sequenced the associated cfDNA (ChIP-Seq) in samples from cancer patients and healthy subjects. The mean observed enrichment of plasma cfCTCF-DNA by ChIP was 180-fold with a mean CTCF recovery of 48%. Background cfDNA, primarily in the form of nucleosomes, was almost completely physically removed (> 99.7%) by ChIP and the remainder was removed bioinformatically. cfCTCF-DNA sequences were partitioned into low or high affinity binding cfCTCF-DNA nucleoproteins in silico, on the basis of binding motif concordance with the consensus CTCF binding motif. We observed that cfCTCF-DNA was quantitatively and qualitatively different in samples from healthy subjects or cancer patients. High affinity cfCTCF-DNA nucleoproteins were present only in the plasma of cancer patients and not in samples from healthy subjects. Testing of small developmental and validation sample cohorts for ChIP-seq peaks for a 14 member panel of high affinity CTCF loci sequences, showed excellent specificity and sensitivity for cancer detection.

Conclusions

Plasma sequence data sets comprising > 99% pure CTCF protected ctDNA can be prepared using a 2-step process involving the physical removal of most background cfDNA in vitro, followed by removal of all remaining background cfDNA in silico. CTCF plasma ChIP-seq represents a novel liquid biopsy method for a novel class of liquid biopsy biomarkers.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s13148-026-02205-x.

Background

CCCTC-binding factor (CTCF) is a highly conserved zinc finger transcription factor (TF) that regulates chromatin structure, insulation and transcriptional control in health and disease [1]. Although CTCF protein expression is ubiquitous, CTCF occupancy of its binding sites (BS) is reported to be tissue specific, cancer specific and to vary with tumor type [1–6]. Tissue and cell line CTCF binding studies by chromatin immunoprecipitation (ChIP) and sequencing of associated DNA (ChIP-Seq) have implicated cancer specific gain and loss of CTCF binding in carcinogenesis, cancer cell proliferation, development of chemoresistance, epithelial-mesenchymal transition and progression to metastatic disease in a variety of cancers including breast, hepatocellular, colorectal, lung and pancreatic cancer [6–13]. CTCF BS occupancy has been proposed as a potential biomarker and a therapeutic target in oncology [5, 10–13].

There is also increasing interest in the analysis of plasma cell free DNA (cfDNA) that circulates bound to CTCF as cell free CTCF-DNA (cfCTCF-DNA) nucleoprotein, as a biomarker for liquid biopsy [14, 15]. The vast majority of circulating plasma nucleoproteins are nucleosomes, in which an approximately 130–200 base pair (bp) cfDNA fragment is bound with high affinity to proteins in a stable histone-DNA complex. However, a small proportion of cfDNA circulates as ultrashort approximately 35–80 bp fragments protected from nuclease digestion by binding to TFs, including CTCF among others [14–16].

The human genome includes tens to hundreds of thousands of CTCF BS loci, with varying binding affinities. A ChIP-seq study of 67 cell lines identified 137,471 non-overlapping CTCF binding peaks, of which approximately 16,000–61,000 are occupied in a single cell type [1, 2, 17–20]. Lower affinity CTCF binding sites are reported to be associated predominantly with transcription regulation and high affinity binding sites with regulation of chromatin structure through topologically associated domains (TADs) [21]. High affinity recombinant CTCF-DNA nucleoproteins are stable with an experimentally measured half-life of up to 22 h [22]. The combination of a long half-life and a large number of binding sites means that a single cell death event may lead to the release of a large number of cfCTCF-DNA nucleoproteins that are sufficiently stable in plasma to provide a detectable level by liquid biopsy.

The presence of cfCTCF-DNA in plasma has been inferred from DNA footprinting and nucleosome fragmentomics analysis of cfDNA sequencing data [14, 15, 23]. Whilst these methods have established the aggregate presence of cfCTCF-DNA and other TF-DNA complexes in plasma, they have provided little or no information on the occupancy of each of the thousands of individual CTCF binding sites represented in the plasma as cfCTCF-DNA. ChIP-seq and related techniques such as ATAC-seq, are powerful tools that have underpinned our understanding of epigenetic control through transcription factor binding to chromatin. ChIP-seq analysis does provide individual TF binding site occupancy data, but direct measurement of endogenous cfCTCF-DNA nucleoprotein by plasma ChIP-Seq has not been reported.

The DNA binding domain of CTCF is an 11 zinc finger protein (ZFP) and ZFP-DNA binding requires zinc ions. Although zinc finger TFs bind strongly to zinc ions with dissociation constants (k_d) in the range of 10⁻⁹–10⁻¹¹ M, EDTA chelates zinc ions some 10,000 times more tightly (k_d 10⁻¹⁶). Exposure of zinc finger TFs to low 0.1-5 mmol/L levels of EDTA, even for a few minutes, may result in significant stripping of zinc ions from ZFPs, disruption of ZFP-DNA binding and dissociation of TF-DNA [24, 25]. We hypothesized that cfCTCF-DNA nucleoproteins might be labile to disruption in plasma which typically contains 2-5 mmol/L EDTA.

TF-DNA ChIP-Seq analysis of cell or tissue sample chromatin typically involves formaldehyde crosslinking to stabilize nucleoprotein structures [26]. Although typical crosslinking protocols involve exposure to formaldehyde for 10 min or more, studies of the reaction kinetics of formaldehyde crosslinking of nucleoproteins report that crosslinking is rapid and complete within 5 s. The remaining time reflects the TF k_on rate and allows for increased binding of low affinity TFs to bind to low occupancy TFBS [27]. As the time required for crosslinking is short (5 s) in relation to high affinity CTCF residence times (up to 22 h), we hypothesized that rapid crosslinking in whole blood might stabilize cfCTCF-DNA nucleoproteins to the ZFP zinc finger depletion effects of EDTA, rendering them amenable to analysis by plasma ChIP-Seq.

We now describe direct isolation and analysis of plasma cfCTCF-DNA nucleoproteins representing high affinity CTCF BS sequences in crosslinked plasma samples from healthy subjects and cancer patients to identify individual CTCF loci sequences selectively present as cfCTCF-DNA in the plasma of cancer, but absent from the plasma of healthy subjects, as potential liquid biopsy oncology biomarkers.

Methods

Crosslinking of cfCTCF-DNA in whole blood

Covalent crosslinking of cf-cfCTCF-DNA nucleoproteins in whole blood was performed by collection of samples in Streck Cell-Free DNA BCT tubes (Streck, La Vista, NE, USA) containing a formaldehyde-releasing agent [28, 29].

Patient samples

Except where otherwise specified, all plasma samples were collected in Streck Cell-Free DNA BCT tubes. Samples were processed according to the manufacturer’s instructions. Separated Streck plasma was stored frozen in cryotubes at -80⁰C until assayed.

Three independent patient sample cohorts were collected. The first cohort was a developmental cohort that included 70 samples collected from 49 treatment naïve patients diagnosed with colorectal cancer (CRC) (n = 10), lung cancer (n = 10), breast cancer (n = 10), prostate cancer (n = 10), liver cancer (n = 5) or cancer of the bile duct (n = 4) purchased from Indivumed (Indivumed GmbH, Hamburg, Germany). Samples collected from 11 healthy donors and 10 patients diagnosed with an inflammatory disorder were purchased from BioIVT (BioIVT LLC Westbury, NY, USA). These 70 samples were analyzed for CTCF protected cfDNA by ChIP-seq using a sample volume of 1 mL.

A second sample cohort including 4 samples from CRC patients was purchased from BioIVT. These samples were use to determine the enrichment and recovery of CTCF protected cfDNA from 1 mL plasma in ChIP-seq experiments.

A third independent validation cohort purchased from Indivumed and BioIVT included 24 samples collected from 10 healthy volunteers, 8 patients diagnosed with CRC and 6 patients diagnosed with lung cancer. These 24 samples were analyzed for CTCF protected cfDNA by ChIP-seq using a sample volume of 3 mL.

Inclusion criteria for healthy subject samples were age ≥ 18 and self-declared as not having any known disease diagnosis. Exclusion criteria for healthy subjects were age < 18. Inclusion criteria for cancer patient samples were age ≥ 18 with a confirmed cancer diagnosis. Exclusion criteria for cancer patient samples were age < 18. Inclusion criteria for inflammatory patient samples were age ≥ 18 with a diagnosis of a chronic inflammatory condition not currently in a flare. Exclusion criteria for inflammatory patient samples were age < 18, current hospitalization or current disease flare.

Informed consent was obtained from all subjects involved in the study and all experiments and methods were performed in accordance with relevant guidelines and regulations. The study followed the principles outlined in the Declaration of Helsinki for all human experimental investigations. All blood specimens were purchased from biobanks where the blood samples are ethically collected with IRB-approved protocols.

Age and gender details of the sample cohort are summarized in Table 1. Individual sample details including age, gender and disease stage are provided in Supplementary Information Table S1–S3.

Table 1.

Patient sample cohorts

	Cancer patients	Healthy subjects	Inflammatory patients
Developmental cohort (1 mL)
Number (male, female)	49 (m26, f23)	11 (m7, f4)	10 (m2, f8)
Age range (mean, median)	43–86 (69, 66)	48–70 (58, 58)	19–58 (34, 36)
Stage I	11
Stage II	12
Stage III	8
Stage IV	18
Colorectal cancer	10
Lung cancer	10
Prostate cancer	10
Breast cancer	10
Liver cancer	5
Bile duct cancer	4
Enrichment cohort (1 mL)
Number (male, female)	4 (m2, f2)
Age range (mean, median)	59–73 (65, 65)
Validation cohort (3 mL)
Number (male, female)	14 (m8, f6)	10 (m3, f7)
Age range (mean, median)	52–82 (67, 68)	19–55 (33, 30)
Stage III	2
Stage IV	11
Stage unknown	1
Colorectal cancer	8
Lung cancer	6

Native and crosslinked MCF7 cell chromatin digest materials

MCF7 breast cancer cells (ATCC) were cultured in MEM medium (Gibco, 31095–029) supplemented with 10% FBS Premium Plus (Gibco, A4766801), NEAA (Gibco, 11140–050), Penicillin–Streptomycin (Gibco, 15070063), and L-Glutamine (Gibco, 25030081). Cells were either used native or crosslinked by fixing for 10 min at room temperature in MEM containing 1% paraformaldehyde before quenching with 125 mM glycine.

Crosslinked or native cell pellets were washed with cold PBS and maintained on ice. Cells were lysed in buffer (10 mM HEPES pH 7.4, 10 mM KCl, 1 mM sodium orthovanadate, 0.05% NP-40 with protease inhibitors) for 20 min on ice. Nuclei were collected by centrifugation, resuspended, and incubated in low-salt buffer (10 mM Tris–HCl pH 7.4, 0.2 mM MgCl₂, 1 mM sodium orthovanadate, 1% Triton X-100 with protease inhibitors) for 15 min on ice. Chromatin pellets were obtained by centrifugation and digested with MNase (0.1 U/µl) for 15 min with agitation. Digestion was stopped by addition of 500 mM EDTA pH 8, followed by 5 min incubation on ice. The supernatant containing fragmented crosslinked or native chromatin was collected after centrifugation.

Nucleosome measurements

Nucleosome measurements were carried out using the automated Nu.Q® H3.1 chemiluminescent immunoassay for nucleosomes containing the histone isoform H3.1 (H3.1-nucleosomes) in 50 µL of plasma as per the manufacturer’s instructions (Belgian Volition SRL, Isnes, Belgium).

Magnetic anti-CTCF-DNA reagent

Anti-CTCF antibody was coated onto magnetic beads (MyOne TosylActivated Dynabeads™) using standard methods. Briefly, antibody was incubated with magnetic beads for 18 h at 37 °C in a rolling bottle to maintain suspension of the beads. The beads were sedimented and resuspended in a blocking buffer comprising phosphate buffered saline (PBS) pH 7.4 containing 1% bovine serum albumin (BSA) for 1 h at 37 °C. The beads were then sedimented again, washed twice and stored in PBS containing 0.1% Tween 20, 1% BSA, and a preservative. Magnetic beads coated with anti-nucleosome antibody (Belgian Volition SRL, Isnes, Belgium) or magnetic beads coated with non-specific mouse IgG (Meridian Life biosciences, Cincinnati, OH, USA) were used as a comparative or control reagent respectively.

Chromatin immunoprecipitation (ChIP) of plasma cfCTCF-DNA

Plasma samples were diluted in a radioimmunoprecipitation assay (RIPA) buffer and anti-CTCF coated magnetic particles were added (30 µg particles/mL plasma). The mixture was incubated for 1 h at room temperature with rolling to maintain suspension of the beads. The beads were then sedimented and washed five times (Supplementary Fig. 1). Anti-nucleosome ChIP was performed in a similar manner using magnetic particles coated with anti-nucleosome antibody.

Western blot

Magnetic anti-CTCF reagent particles were exposed to 1.6 mL plasma suspended in a denaturing 1% sodium dodecyl sulfate (SDS) buffer. The denatured protein was analyzed by western blot using anti-CTCF antibody and a labeled anti-IgG detection antibody (Abcam, VeriBlot (HRP) ab131366).

DNA sequencing

CTCF associated cfDNA isolated by ChIP from 1 ml Streck plasma was extracted from magnetic beads using a QIAamp DSP Circulating Nucleic Acid Kit (QIAGEN, Hilden, Germany), according to the manufacturer’s instructions. The extracted cfDNA was amplified to produce a single strand library for sequencing using a SRSLY NGS Library Prep Kit (Claret Bioscience, Santa Cruz, CA, USA), according to the manufacturer’s instructions, and sequenced at a coverage of approximately 30X using an Illumina NovaSeq Sequencing System.

The 70 developmental plasma samples were analyzed to identify cfCTCF-DNA in a single batch. The 4 CRC patient samples analyzed for CTCF sequence panel validation and refinement were sequenced in a second single batch. The 24 validation cohort samples were sequenced in a third single batch.

Processing of sequencing reads and fragment size selection

Sequenced reads were processed using the Nextflow [30] nf-core [31] ATAC-seq pipeline v2.1.2 (https://doi.org/10.5281/zenodo.2634132), including adapter trimming (Trim Galore v.0.6.7, using 20 bp trimming to accommodate NovaSeq chemistry), alignment (BWA-MEM v.0.7.17-r1188) [32], blacklist filtering (https://github.com/Boyle-Lab/Blacklist/blob/master/lists/hg38-blacklist.v2.bed.gz), and low-quality read filtering (to filter out duplicate reads, multi-mapped reads and split (secondary and supplementary) reads, and reads that contained more than 4 mismatches or were improperly paired). Quality metrics on the sequencing reads were monitored using MultiQC [33]. Fragment size frequency diagrams were produced for filtered sequences using the CollectInsertSizeMetrics tool from Picard (v3.0.0), and plotted using Multi-QC. All sequence alignments and subsequent analyses used human genome assembly GRCh38.

The resulting alignment BAM files were used to create subsets of different fragment sizes (for example; 35‒80 bp, 135‒155 bp and 156‒180 bp, inclusive of the boundaries) using the Template Length (TLEN) of the Sequence Alignment/Map with SAMtools [34]. Any non-aligned reads were discarded, as were those with mapping quality below 30 (MAPQ -q 30) or if the reads failed one of the mapping criteria using the bitwise flags in SAM column 2 (-F 3852 = 4 + 8 + 256 + 512 + 1024 + 2048).

Read coverage (the number of fragments found to cover a specific gene locus) was calculated using a bin size of 1 bp. Read coverage was normalized to the total number of reads mapped to the human genome with the reads per genome coverage (RPGC) using the deepTools bamCoverage. Coverage profile plots were generated for each fragment size using deepTools plotProfile [35].

CTCF ChIP-Seq peak calling

ChIP-Seq peak calling was performed on BAM files of the 35-80 bp fragments using MACS3 (v3.03) narrow peaks [36] (the parameters used were -t, -g 2.7e + 9, -m 5 50, -f BAMPE). The peak files were used to detect transcription factor binding sites using the consensus probability matrix for CTCF described in the HOMER database [37, 38]. No correction for background binding to non-specific IgG was used in peak calling (parameter -c). Peaks were picked using False Discovery Rate (FDR) adjusted p-values between ≤ 0.05 to ≤ 0.00001 [39].

CTCF binding affinity

CTCF binds to a consensus 20 base DNA binding motif and the affinity of binding at particular loci can be estimated by its similarity to the consensus motif [4, 21, 37, 38]. CTCF binding site loci identified by plasma ChIP-seq were interrogated for a parameter of binding affinity based on a moving score for the conformity of the peak location sequence with the position probability matrix (PPM) reported for the 20 base consensus binding motif [4]. Starting at a genomic location (n) at the first base occurring at each end of the peak cfDNA fragment sequences observed, the affinity of the location sequence for CTCF binding was calculated as the sum of a score awarded to each of the 20 bases starting at that position. The score awarded corresponded to the probability for that base reported for the probability consensus sequence (e.g. if the location sequence included an adenine at a certain location where the probability for an A in the probability consensus sequence was 0.7, then the peak location sequence was awarded a score of 0.7 for that base). The parameter used for CTCF location and binding affinity was the sum of the scores for each of the 20 nucleotide bases. The process was repeated starting at genomic location (n + 1), (n + 2), (n + 3) etc. to cover the broader cfDNA fragment location sequence, and the highest score achieved in either direction was used to identify the CTCF binding motif location and its CTCF binding affinity. An average score of ≥ 0.5 for each base, or a total score of ≥ 10 for the 20 base pairs in this motif, was considered high affinity.

cfCTCF-DNA ChIP-seq differential peak analysis

We identified CTCF ChIP-seq peaks present in plasma samples but not in cancer patients and healthy subjects, using increasingly strict FDR p-value cutoffs for CTCF protected cfDNA fragments in the size range 35–80 bp that included a CTCF binding site sequence [39]. cfDNA sequences representing CTCF binding sites represented as short cfDNA in cancer samples, but not in healthy or inflammatory samples, were identified using the DiffBind package from BioConductor running on R using DESeq2 in the normalization step [31–33, 40, 41]. FRiP scores (Fraction of Reads in Peaks) were calculated for all samples according to the ENCODE guidelines [42].

UniBind analysis of peaks from validation sample cohort

We constructed a reusable autosomal CTCF reference from UniBind [43] by downloading the Homo sapiens TFBS archive, extracting 556 CTCF ChIP-seq sample files, and merging 561,898 unique strand-specific CTCF motifs contained in one or more of those samples (after excluding chrX and chrY). Motifs were grouped using a permissive ± 3 bp overlap rule, with final coordinates defined by the union of original intervals within each group and the motif score set to the maximum Position Specific Scoring Matrix (PSSM) score, yielding 540,924 merged autosomal UniBind CTCF sites. We then selected the subset that overlapped MACS3 peak intervals from any validation set cfChIP-seq sample (q-value cutoff 0.05) and annotated each peak with its maximum overlapping MACS3 -log10(qvalue). In addition to these 10,753 UniBind-derived peaks regions, we merged all remaining peaks using the same ± 3 bp rule to result in 1,784 non-UniBind peaks. The final 12,537 peak regions were used for all downstream analysis. As an independent control, we randomly selected a set of 12,537 UniBind CTCF motif regions without respect to peak overlap.

For rank-based cfChIP-seq genomic analysis, fragment counts were obtained from indexed BAM files (including only reads with mapping quality at least 30 and TLEN fragment length 35–80 bp). Within each sample, counts were converted to ranks and visualized as rank-based heatmaps. To derive sample-level summaries for fractions overlapping ubiquitous CTCF sites and fractions overlapping RepeatMasker, we selected the top 1,000 ranked sites per sample and calculated fractions overlapping from these sites only. Ubiquitous CTCF sites were defined as those overlapped in at least 500 UniBind CTCF ChIP-seq samples. RepeatMasker was obtained from LOLA [44].

Results

Stability of exogenous MCF7 CTCF-cfDNA spiked into EDTA plasma

To determine the stability of native and formaldehyde cross-linked CTCF-DNA nucleoproteins in EDTA plasma, we spiked native and crosslinked MCF7 cell chromatin digests into healthy native EDTA plasma containing no formaldehyde releasing agent. CTCF-DNA nucleoproteins were isolated from the spiked plasmas by ChIP and analyzed for the fragment size profile of the isolated cfDNA sequence data. We observed a large peak in the ultrashort 35–80 bp size range reported for CTCF-associated DNA fragments [14] in the ChIP isolate of cross-linked chromatin digest spiked plasma, but no peak for native chromatin (Fig. 1a). As EDTA is reported to disrupt ZFP-DNA binding by sequestration of zinc ions, we concluded that native exogenous cfCTCF-DNA complexes are labile to dissociation in EDTA plasma and that dissociation can be prevented by formaldehyde cross-linking [24, 25].

Fig. 1.

a size frequency of diagram of DNA fragments isolated by CTCF ChIP from native EDTA plasma spiked with crosslinked (red) or native (blue) exogenous digested MCF7 cell chromatin b Western Blot analysis of endogenous CTCF isolated by CTCF ChIP from four Streck CRC samples;

Characterization of endogenous plasma protein isolated by CTCF ChIP

CTCF ChIP was performed on 1 mL Streck plasma samples collected from four colorectal cancer patients. The ChIP isolates were analyzed for the presence of CTCF protein by western blot. A band at approximately 140 kD, characteristic of CTCF protein [45], was present in the isolates of all four samples using specific anti-CTCF antibody coated magnetic beads (Fig. 1b). We concluded that the CTCF ChIP isolates derived from the 4 CRC samples contained CTCF.

Characterization of endogenous plasma cfDNA isolated by CTCF ChIP

CTCF ChIP isolated cfDNA would be expected to include ultrashort 35–80 bp fragments that include a CTCF binding motif and flanking sequences but not include nucleosome size cfDNA fragments including the same sequences (as nucleosomes are removed by ChIP). We therefore investigated the size range of cfDNA fragments present in cfDNA extracted from 4 CRC whole plasma samples (1 mL) as well as from CTCF ChIP isolates and nucleosome ChIP isolates from the same samples. The cfDNA fragment size profile of the whole plasma samples had a peak at approximately 170 bp characteristic of nucleosome associated cfDNA as well as a shoulder of smaller fragments of less than 100 bp in length (Fig. 2a). In order to determine whether the shorter cfDNA fragments present were of non-nucleosomal origin or were nucleosome derived from nicked nucleosomes containing one or more single or double stranded breaks, we investigated the cfDNA profile on nucleosome ChIP isolates from the same 4 samples using an anti-nucleosome antibody. The observed cfDNA size profiles observed for all 4 nucleosome ChIP isolates were similar to the whole plasma size profiles indicating that the ultrashort cfDNA was in each case comprised primarily of nucleosome associated cfDNA (Fig. 2a). The whole plasma cfDNA sequencing data was then analyzed for cfDNA fragments that included one of 9,780 previously published ubiquitous CTCF binding sites [46, 47]. The results for whole plasma sample cfDNA included 2 peaks: one peak at approximately 170 bp corresponding to nucleosome protected cfDNA and another peak at approximately 70 bp (Fig. 2b). This shorter peak is consistent with the approximately 35–80 bp size range reported previously for TF protected cfDNA obtained indirectly by DNA footprinting [14, 15] and also consistent with the size of CTCF protected DNA fragments we and others have observed directly for nuclease digests of cell line chromatin [48].

Fig. 2.

cfDNA fragment size profiles for endogenous cfDNA fragments from a Streck CRC plasma sample: a all cfDNA extracted from whole plasma (black) or a nucleosome ChIP isolate from plasma (blue); b cfDNA extracted from whole plasma (blue) or a CTCF ChIP isolate from plasma (red) including one of 9,780 published ubiquitous CTCF binding sites isolated by CTCF ChIP from a Streck CRC plasma sample

We next analyzed the size profile of cfDNA fragments isolated by CTCF ChIP from the same 4 plasma samples (1 mL) that included one of the same 9,780 previously published CTCF binding sites [47]. These isolates contained only a single peak at 70 bp indicating that CTCF protected cfDNA was retained by CTCF ChIP but nucleosome protected cfDNA fragments bearing the same sequences were absent (Fig. 2b).

We analyzed further 1 mL healthy and cancer samples for coverage of the same 9,780 CTCF binding sites by cfDNA fragments in 3 size ranges of 35-80 bp, 135-155 bp and 156-180 bp [47]. CTCF ChIP isolated cfDNA fragments in the 35-80 bp size range, but not in the 135–180 bp ranges, were enriched in coverage plots for the 9,780 sequences (Fig. 3a). No enrichment for any fragment size was observed for non-specific ChIP isolates obtained using magnetic particles coated with non-specific IgG (Fig. 3b).

Fig. 3.

Analysis of cfDNA isolated from plasma by anti-CTCF or anti-nucleosome ChIP. a CTCF coverage by endogenous cfDNA fragments sized 35–80 bp, 135–155 bp or 156–180 bp isolated from a CRC patient sample by CTCF ChIP using anti-CTCF antibody coated magnetic particles; b as per (a) but using non-specific IgG coated magnetic particles; c V-plot showing the size and location of a CTCF binding site motif within cfDNA fragments isolated by CTCF ChIP from a CRC plasma sample (intensity range 0–5000); d as per (c) but for cfDNA fragments isolated by ChIP using anti-nucleosome antibody coated magnetic particles (intensity range 0–120) and e as per c but for background binding of cfDNA fragments bound to non-specific IgG coated magnetic particles (intensity range 0–120)

V-plots of the sequence data showed that the CTCF associated cfDNA comprised predominantly fragments in the ultrashort 60-80 bp range and that the CTCF binding motif was non-randomly located approximately centrally within the ultrashort cfDNA fragments (Fig. 3c) [49]. In contrast, V-plots of the 9,780 CTCF binding site sequences isolated by ChIP using an anti-nucleosome antibody were 160-170 bp in length and the CTCF binding motif was randomly positioned within the fragments (Fig. 3d). V-plots of the same sequences isolated using non-specific IgG coated magnetic particles showed no enrichment of the 9,780 CTCF binding site sequences at any fragment size or location within a fragment (Fig. 3e).

Endogenous cfDNA fragments isolated by CTCF ChIP from plasma were: (1) bound by anti-CTCF coated magnetic beads but not bound by non-specific IgG coated beads, (2) were of the sub-nucleosomal size range expected for TF associated DNA, (3) were enriched for published genomic CTCF binding site sequences; and (4) included a centrally located CTCF binding motif. We concluded that CTCF ChIP isolates from Streck plasma samples were enriched for endogenous crosslinked cfCTCF-DNA nucleoproteins.

Enrichment of plasma cfCTCF-DNA by ChIP

Partitioning of cfDNA by ChIP involves isolation of cfCTCF-DNA and removal of nucleosomes and other non-CTCF bound cfDNA. As the vast majority of cfDNA is nucleosome-bound, the ChIP isolates retained only a small proportion of the total cfDNA present in the whole sample, but containing a higher proportion of CTCF-bound cfDNA subtypes. We estimated the degree of enrichment of cfCTCF-DNA in CTCF ChIP isolates as the frequency of ultrashort cfDNA fragments that included one of 9,780 published CTCF binding sites in the cfDNA extract from a whole plasma sample (1 mL) compared to the frequency of the same fragments in a CTCF ChIP isolate from the same sample (1 mL) [47]. We sequenced whole plasma cfDNA, and CTCF ChIP isolated cfDNA, from Streck samples obtained from 4 CRC patients. Patient details and ChIP-seq FRiP scores are listed in Supplementary Table 2. The frequency of ultrashort cfDNA fragments that included a published CTCF binding site [46, 47] and the enrichment of CTCF associated ultrashort cfDNA fragments by ChIP were calculated as:

The mean number of aligned reads present in the 4 CRC whole plasma samples (1 mL) was 73.5 million. Of these reads, 11,329 reads were ≤ 100 bp in length and included one of the 9780 sequence matches. The mean number of total aligned reads present in the CTCF ChIP isolates from 1 mL of plasma was 197,000 of which 5,470 reads were ≤ 100 bp in length and included one of the 9780 sequences (Table 2). We concluded that 99.73% of plasma cfDNA was removed by ChIP, but that the remaining 0.27% of plasma cfDNA isolated by CTCF ChIP retained 48.3% of the ultrashort 9780 CTCF binding site reads present in whole plasma.

Table 2.

Frequencies and enrichment of ultrashort CTCF bound cfDNA fragments including one of 9,780 published CTCF binding site sequences in whole plasma and in plasma ChIP isolates from 4 CRC patients

	Sample 1	Sample 2	Sample 3	Sample 4	Mean
Whole plasma
Total aligned reads	74,739,072	83,265,426	74,413,111	61,387,310	73,451,230
≤ 100 bp CTCF reads	12,319	10,529	11,653	10,814	11,329
Frequency CTCF reads	0.016%	0.013%	0.016%	0.018%	0.015%
CTCF ChIP isolate
Total aligned reads	50,191	370,220	277,582	88,682	196,669
≤ 100 bp CTCF reads	1,611	12,439	6,023	1,806	5,470
Frequency CTCF reads	3.21%	3.36%	2.17%	2.04%	2.78%
Fold enrichment	195	266	139	116	180

The mean frequency of ChIP targeted ultrashort cfDNA fragments including one of the 9,780 sequence matches in the 4 ChIP isolates was 2.78% compared with a mean of, in the whole plasma samples. This equates to a mean CTCF associated cfDNA enrichment of 180-fold (Table 2). As the number of CTCF BS reported to be occupied in a typical cell is greater than 9,780, the actual proportion of CTCF associated cfDNA may have been greater than 2.78%. A frequency diagram for the size of cfDNA fragments bearing one of the 9,780 CTCF binding site sequences in a CTCF ChIP isolate is shown in Fig. 2b. We concluded that CTCF ChIP isolates from Streck plasma were highly enriched for endogenous cfCTCF-DNA.

cfCTCF-DNA ChIP-seq peaks in 1 mL developmental sample cohort

We first investigated the enrichment of ultrashort 35–80 bp cfDNA fragments bearing one of 9,780 published CTCF binding site loci sequences as an indicator of successful isolation of cfCTCF-DNA in the 49 cancer samples and 21 control samples of the developmental cohort (Table 1). Enrichment, as exemplified in Fig. 3a, was observed for all cancer and control samples tested.

We characterized a plasma cfDNA fragment as CTCF bound if it was bound by an anti-CTCF antibody, was sub-nucleosomal in size and included an approximately centrally located CTCF binding motif sequence. The motif and upstream and downstream flanking sequences were used to identify the locus of the binding site in the genome.

We investigated CTCF ChIP-seq peaks represented as cfCTCF-DNA in the sequence data for 1 mL plasma samples using bioinformatics as follows:

1. Aligned cfDNA fragment sequences in the size range 35–80 bp were identified and fragments of other sizes were excluded from further analysis. Exclusion of larger nucleosome size cfDNA fragment sequences removed sequence data pertaining to non-specific binding of nucleosomes to the magnetic particles used in ChIP.

2. CTCF ChIP-seq peaks in the 35–80 bp cfDNA fragment sequence data were identified bioinformatically using the HOMER software package, with FDR p-values ≤ 0.05 and ≤ 0.01 (equivalent to a random event rate of 1:20 to 1:100).

3. All ChIP-seq peaks identified were interrogated for the affinity of the CTCF-DNA binding motif sequence present, based on a moving score for the conformity of the peak location sequence with the 20 base CTCF binding motif PPM. ChIP-seq peaks related to cfCTCF-DNA including a binding motif sequence with a score < 10 were removed. This step removed all cfDNA sequence data that did not include a CTCF binding site and additionally removed all peaks related to low affinity CTCF binding sites. Interestingly, this step also removed peaks with repetitive motif sequences leaving binding loci of high motif sequence complexity.

We observed a total of 15,791, 14,259 and 9,988 ChIP-seq peaks cumulatively present among all 70 plasma samples in the 1 mL developmental cohort at FDR p ≤ 0.05, ≤ 0.01 and ≤ 0.00001 respectively. The proportion of these ChIP-seq loci present in cancer samples was high and increased with the use of increasingly stringent FDR p-values (96.4%, 99.3% and 99.9% respectively at FDR p ≤ 0.05, ≤ 0.01 and ≤ 0.00001). However, the proportion present in healthy or inflammatory samples was lower and decreased with increasingly stringent FDR p-values (40.9%, 27.9% and 15.2% respectively for healthy samples and 4.2%, 1.2% and 0.1% respectively for inflammatory samples). FRiP scores were > 0.01 for 69 of 70 plasma samples. Mean FRiP scores were 0.033, 0.063 and 0.078 for 1 mL healthy, inflammatory and cancer samples respectively (range 0.007–0.221). Individual FRiP scores are listed for each sample in Supplemental Table 1.

The data included a large number of potential candidate ChIP-seq peaks representing cfCTCF-DNA loci selectively present in the plasma of cancer patients. In view of this, we selected a subset of 265 ChIP-seq peaks that were present in at least 10 of the 49 cancer samples tested, with no false positive results among the 11 healthy and 10 inflammatory control samples (1 mL) tested. From these 265 loci we identified 14 loci selected empirically as representing the smallest panel that gave a positive result for all 49 cancer samples, including 11 stage I and 12 stage II cancers, with no false positive results among control samples. The minimal 14 CTCF binding site sequences were used as an example panel for testing in a validation sample cohort (Table 3).

Table 3.

Fourteen CTCF Binding site loci selected empirically as an example panel for testing in a validation sample cohort

Locus	Nearest gene	Chromosome	Start location	End location
1	COL4A3	chr2	227238419	227238516
2	NKAIN1	chr1	31162814	31162900
3	NEDD4L	chr18	58180897	58180987
4	APBB2	chr4	40908810	40908904
5	DYNC1I1	chr7	95916718	95916810
6	GSTM4	chr1	109640350	109640454
7	NRP2	chr2	205766899	205766986
8	INPP4A	chr2	98569347	98569426
9	HBEGF	chr5	140351878	140351960
10	EPO	chr7	100725425	100725514
11	SVIL	chr10	29590831	29590908
12	LINC02389	chr12	64891019	64891105
13	ABR	chr17	1187325	1187419
14	PKIA	chr8	78393565	78393649

cfCTCF-DNA ChIP-seq peaks in 3 mL blinded validation sample cohort

The 14 pre-selected CTCF binding site loci sequences selected using the developmental cohort, were tested in a blinded validation cohort of plasma samples collected from 10 healthy volunteers and 14 treatment naïve stage III/IV patients diagnosed with a colorectal (n = 8) or lung (n = 6) cancer (Table 1). Details of the individual subjects are listed in Supplemental Table 3.

As the number of aligned reads obtained for CTCF ChIP isolates from a 1 mL plasma sample was low, we increased the DNA input for sequencing by using an increased sample volume of 3 mL for the validation cohort. The mean numbers of aligned sequence reads obtained for ChIP isolates from the 3 mL samples were 456,458 (range 80,548–1,848,557) from healthy subjects and 877,664 (range 143,114–4,102,451) from cancer patients. FRiP scores obtained for 3 mL cancer plasma samples were higher than observed for 1 mL samples. Mean FRiP scores were 0.041 and 0.209 for healthy and cancer samples respectively (range 0.0097–0.543). Individual FRiP scores are listed for each sample in Supplemental Table 3. The number of reads observed increased with increasing levels of cell free circulating chromatin, measured as the level of H3.1-nucleosomes, but was similar for healthy or cancer samples with similar H3.1-nucleosome levels (Supplementary Fig. 2).

We observed an enrichment of ultrashort 35–80 bp cfDNA fragments bearing one of 9,780 published CTCF binding site loci sequences (as exemplified in the enrichment plot in Fig. 3a) for all 14 cancer samples and for 9 of the 10 control samples tested. The normalized coverage of enrichment plots observed was higher for cancer samples than healthy samples (Fig. 4a). The healthy sample for which no enrichment was observed (sample 3) was excluded from the validation results.

Fig. 4.

a normalized coverage of enrichment plots for 35–80 bp cfDNA fragments in ChIP isolates from healthy and cancer samples; b the effect of FDR p-value used in ChIP-seq peak calculation on the proportion of total ChIP-seq peaks observed that occur only in cancer samples; c the effect of FDR p-value used in ChIP-seq peak calculation on the proportion of total peaks observed in healthy and cancer samples; d similar to (c), but showing loss of peaks at highly stringent FDR on a logarithmic scale

ChIP-seq analysis of the validation cohort resulted in a positive peak for at least one of the pre-identified panel sequences for 13 of the 14 cancer patients tested with a complete absence of any peak for any healthy control subject, representing 93% sensitivity and 100% specificity in this small sample set. Detailed results for the validation cohort are shown in Table 4.

Table 4.

Results for a pre-determined 14-member CTCF locus panel in a blinded validation experiment. A black square represents a positive cfCTCF-DNA ChIP-seq peak result for the plasma sample at the denoted CTCF binding site locus. A white square denotes the absence of a ChIP-seq peak at the denoted locus. The panel of 14 CTCF loci are numbered as in Table 3. The validation cohort of 24 samples are presented in ascending order of the number of loci positive and numbered 1–24 in the blinded order in which they were studied. Healthy sample 3 was excluded from validation analysis of sensitivity and specificity as no enrichment of ultrashort cfCTCF-DNA (as illustrated in Fig. 3a) was observed on ChIP-seq analysis (although sample 3 would be classified as healthy if included)

Quantitative investigation of plasma cfCTCF-DNA ChIP-seq peaks

We next investigated the numbers of distinct cfCTCF-DNA ChIP-seq peaks observed for plasma samples from healthy subjects and cancer patients. A surprising but consistent feature of the data was a universal finding of ChIP-seq peaks representing high affinity CTCF binding site loci sequences in cancer samples, but a near complete absence of high affinity peaks from the plasma of healthy subjects.

A total of 12,219 and 11,795 distinct ChIP-seq peaks were observed cumulatively present among all 14 cancer samples in the 3 mL validation cohort at FDR p ≤ 0.05 and ≤ 0.01 respectively. This represents 100% and 98.8% respectively of the total ChIP-seq peaks observed among all 23 samples (both healthy and cancer samples combined). In contrast, the cumulative numbers of peaks observed in all healthy samples combined were 602 and 146 at the same FDR p-values representing 0.049 (4.9%) of total peaks at FDR p ≤ 0.05, and 0.012 (1.2%) of total peaks at FDR p ≤ 0.01. As the number of peaks in healthy samples was close to that expected to be generated randomly, we concluded that the numbers of distinct CTCF binding site loci represented as cfCTCF-DNA in plasma samples from healthy subjects was small and may represent random events or bioinformatic artifacts.

As genuine and artifactual ChIP-seq peaks may respond differently to calculation using different FDR p-values, we calculated the numbers and proportions of distinct peaks present in CTCF ChIP isolates from healthy and cancer samples using a range of FDR p-values. Increasing the stringency of FDR p-value used to calculate the number of peaks present in cancer samples, from ≤ 0.05 to ≤ 0.00001, led to a reduction of 29.9% in the absolute number of peaks observed, but an increase in the proportion of peaks present only in cancer samples (but not in any healthy samples) from 95.1 to 99.9% (Table 5 and Fig. 4b). In contrast, both the number and proportion of total ChIP-seq peaks calculated for CTCF ChIP isolates from healthy samples decreased markedly from 602 to 5, or 4.9% to 0.06% of total peaks. We concluded that high affinity ChIP-seq peaks observed for cancer samples are robust to calculation at highly stringent FDR p-values, but those observed for healthy subjects are less robust and their number tends to zero with increasing stringency of FDR (Fig. 4c and d), and that peaks calculated at stringent FDR p-values represent pure tumor derived CTCF protected ctDNA.

Table 5.

The number and proportion (%) of ChIP-seq peaks (distinct occupied CTCF binding site loci) observed for CTCF ChIP isolates from 3 mL healthy and cancer patient samples calculated with decreasing p-value for FDR

Number of ChIP-seq peaks	False discovery rate adjusted p-value
	≤ 0.05	≤ 0.01	≤ 0.005	≤ 0.001	≤ 0.0001	≤ 0.00001
Peaks all samples	12,219	11,942	11,580	10,415	9,865	8,562
Peaks in any of 14 cancer samples	12,219 (100%)	11,931 (99.9%)	11,571 (99.9%)	10,407 (99.9%)	9,865 (100%)	8,562 (100%)
Peaks only in cancer samples	11,617 (95.1%)	11,795 (98.8%)	11,436 (98.8%)	10,307 (99.0%)	9,828 (99.6%)	8,557 99.9%)
Peaks in any of 9 healthy samples	602 (4.9%)	146 (1.2%)	144 (1.2%)	108 (1.0%)	37 (0.4%)	5 (0.06%)

The mean number of ChIP-seq peaks in an individual CRC or lung cancer sample was approximately 25% and 30% of total peaks respectively. The mean proportions of peaks present in an individual lung cancer or CRC sample calculated at varying FDR are shown in Supplementary Fig. 3. Whilst ChIP-seq peaks observed for cancer samples were highly robust to recalculation at low FDR p-values, those observed for healthy samples at FDR p ≤ 0.05 were almost entirely absent at low FDR p-values. We concluded: (1) that calculation of ChIP-seq peaks using a strict FDR p-value eliminates plasma CTCF ChIP-seq peaks representing cfDNA derived from healthy cells; and (2) th

at DNA sequence data sets relating to CTCF ChIP-seq peaks calculated for cancer samples at strict FDR p-values therefore represent sequences associated with > 99% pure ctDNA (because they occur only in samples from cancer patients).

The specificity of ChIP-seq peaks we observed for cancer was remarkable. At FDR p ≤ 0.00001 we identified 8,562 peaks that were present for any of the 14 cancer samples, with a total of 5 peaks in any of the 9 healthy samples. This represents a non-tumor background signal rate of 6.5 × 10^–5 (5 background signal events among a possible 77,058 events). A visual representation of the first 130 of 11,795 rows of a table where each row represents a ChIP-seq peak at FDR p ≤ 0.01 and each column represents a healthy, CRC or lung cancer sample is shown in Supplementary Fig. 4.

MultiQC data for the validation sample cohort is shown in Supplementary Information.

Qualitative investigation of plasma cfCTCF-DNA ChIP-seq peaks

To provide more context for the observed quantitative differences between healthy and cancer samples in the validation cohort dataset, we performed a comprehensive genomic analysis based on a large database of CTCF ChIP-seq binding sites from 556 diverse cell types (UniBind, [43]) and a published analysis of cancer-specific CTCF ChIP-seq sites [6]. While UniBind contains 540,294 CTCF motifs overlapped in at least one CTCF ChIP-seq sample, the bulk of these are overlapped by a tiny number of sample types and have low motif matching (Position-Specific Scoring Matrix) scores, as shown in Fig. 5a, right-hand side (“Random UniBind CTCF Sites”). In contrast to random UniBind CTCF sites, the majority of the cfChIP-seq peaks overlapping UniBind sites (10,753) were nearly ubiquitous across UniBind samples and had high-scoring CTCF motif scores, indicating high-affinity binding (Fig. 5b). An additional 1,784 cfChIP-seq peaks that did not overlap a UniBind CTCF motif were also included in the “cfChIP-seq Peak Sites” heatmap, and these sites tended to overlap repetitive sequences from the RepeatMasker database at much higher rates (Fig. 5b). These motif-less, repetitive peak sites had much lower significance based on MACS3 -log10(q-value).

Fig. 5.

Genomic properties of CTCF cfChIP-seq peak regions. a Left panel shows 12,537 peaks called in any cfChIP-seq sample, while right panel shows an equal number of randomly sampled CTCF binding sites from the a total of 540,924 CTCF sites from the UniBind database. Heatmap rows correspond to genomic sites, and heatmap columns correspond to 556 CTCF ChIP-seq experiments from UniBind, with blue color corresponding to an overlapping peak in UniBind. b Genomic features of genomic sites, including 14 cancer marker regions from this study and published cancer-specific CTCF binding sites [6]. CTCF motif (PSSM) score, # samples overlapped, and CTCF overlap are from UniBind database, and -log10(MACS3 q-value) corresponds to most significant q-value obtained for the region for any cfChIP-seq sample (a value of 0 in the right-hand column indicates no peak called). c Heatmaps show rank position (from 1 to 12,537) of all genomic sites within each cfChIP-seq sample, based on number of reads overlapping the site (sites within the left and right heatmaps are ranked independently). Annotation columns show sample-specific metadata (Diagnosis red for cancer and blue for healthy, and H3.1-nucleosome plasma concentration in ng/mL, sequencing QC data (num fragments, FRiP, and deepTools fingerprint value for the 95th percentile of genomic regions). Fraction RepeatMasker and Fraction Ubiquitous CTCF are determined by taking the top 1000-ranked genomic sites, and quantifying the percentage overlapping Repeat Masker or Ubiquitous cellular CTCF sites (those with peaks called in > 500 UniBind CTCF ChIP-seq samples). d Heatmaps show peaks called by MACS3 in each cfChIP-seq sample at a q-value of 0.05. Metadata and sample ordering are duplicated from the panel above. e Comparison of the Fraction RepeatMasker and Fraction Ubiquitous for each sample

The genomic context of cfChIP-seq peaks provided additional biological insights into our 14-marker panel sites, which tended to occur within the most ubiquitous UniBind CTCF sites (Fig. 5b). Interestingly, only one cancer-specific CTCF site from a previous study [6] overlapped a cfChIP-seq peak (Fig. 5b, left); most occurred within sites that were restricted to small, specific subsets of UniBind samples (Fig. 5b, right).

In order to provide a holistic view of genomic peak context and cfChIP-seq sample features, we plotted rank-normalized read counts (Fig. 5c) and MACS3 peak locations (Fig. 5d) for each sample, clustering by rank-normalized read counts. This analysis showed that the healthy samples tended to have peaks and higher read counts in the repetitive sites not overlapping UniBind CTCF motifs, whereas the cancer samples tended to have the converse, with most sites in the UniBind ubiquitous CTCF sites. We quantified this by calculating the fraction of the 1000 top-ranked sites for each sample overlapping either RepeatMasked (‘Frac. RepeatMasker’ annotation column) or ubiquitous UniBind sites (‘Frac. Ubiq CTCF’). These two features were inversely correlated, as quantified in Fig. 5e. The fraction overlapping ubiquitous CTCF sites values were also correlated to H3.1 levels, the MACS3 Fraction of Reads in Peaks (FRiP) score, and to a lesser extent the total number of 35–80 bp fragments (Fig. 5c and d), suggesting these properties are linked at either the biological or technical level. MultiQC data for the validation sample cohort is shown in Supplementary Information.

We concluded that the presence of high affinity ChIP-seq peaks in the plasma of cancer patients, but not healthy subjects, is therefore likely to be related more to their absence from healthy plasma, rather than cancer specific occupation of binding sites only in cancer cells. In contrast, CTCF binding site loci identified in samples from healthy subjects had little overlap with ubiquitous UniBind sites and intersected predominantly with sequences included in the ENCODE blacklist and other repetitive elements that are likely to be poorly mappable (Fig. 5c–e) [50]. These findings are congruent with the finding that ChIP-seq peaks in healthy subjects are removed by filtering for binding strength score and by the use of stringent FDR adjusted p-values for their calculation.

Discussion

Immunochemical partitioning of plasma nucleoproteins to enrich circulating TF protected cfDNA has not previously been reported. For the first time, we isolated and analyzed endogenous plasma cfCTCF-DNA nucleoproteins by ChIP-seq. Although only a small fraction of total cfDNA is CTCF associated, we were able to isolate plasma cfCTCF-DNA by ChIP with a mean enrichment of 180-fold (18,000%) and a mean recovery of 48% of CTCF bound cfDNA fragments. Background cfDNA, primarily in the form of nucleosomes, was almost completely removed (> 99.7%).

Surprisingly, CTCF ChIP-seq studies demonstrated a complete, or near complete, absence of high affinity cfCTCF-DNA nucleoproteins in the plasma of healthy subjects, indicating that the thousands of distinct high affinity cfCTCF-DNA nucleoproteins present in the plasma of cancer patients are tumor derived. This, in turn, indicates that high affinity CTCF ChIP-seq results obtained for cancer patients, represent pure ctDNA sequence data sets. The pure ctDNA data sets include only a minority of, not all, ctDNA fragment sequences present in the plasma, but contain no sequence data related to cfDNA fragments originating from healthy cells.

Plasma CTCF ChIP-seq therefore opens the possibility of a 2-step method for preparing pure ctDNA data sets for cancer patients. In the first step, cfCTCF-DNA is physically isolated in vitro with removal of > 99% of background non-tumor derived cfDNA. In the second step, remaining background cfDNA, including remaining nucleosome protected cfDNA and low affinity CTCF protected cfDNA, is removed in silico by bioinformatics to produce a data set of CTCF ChIP-seq sequences representing pure ctDNA fragments in the plasma sample.

To our knowledge, the observation of an absence of high affinity cfCTCF-DNA nucleoproteins in the plasma of healthy subjects is a novel finding for which the underlying biology is unknown. However, all cells are capable of depletion of CTCF bound to high affinity binding sites associated with the 10,000–20,000 TADs in the nucleus as this occurs during mitosis [51]. Possible speculations regarding the underlying biology include that circulating nucleoproteins in healthy subjects may derive primarily from neutrophils or other leukocytes by NETosis whereby unwound chromatin is released from the cell into the circulation and that unwinding of TADs may similarly be associated with this process.

The aggregate presence of cfCTCF-DNA nucleoproteins in plasma has been demonstrated previously through direct DNA footprinting of ultrashort CTCF bound cfDNA fragments and indirectly by footprinting of nucleosome bound cfDNA. However, detailed analysis of the occupancy of individual CTCF binding sites by footprinting of whole genome sequencing data of whole plasma sample cfDNA extracts has not been reported.

We investigated the proportion of sub-100bp cfDNA that is nucleosome bound by comparing the size frequency profile of cfDNA extracted from whole plasma with cfDNA isolated from plasma by binding to an anti-nucleosome antibody. The results in Fig. 2a demonstrate total cfDNA includes a substantial proportion of ultrashort sub-100bp cfDNA and that the vast majority of this ultrashort cfDNA is derived from nicked nucleosomes. The frequency of nicked nucleosome derived cfDNA fragments carrying CTCF binding motif sequences will likely exceed the frequency of low abundance true CTCF protected ultrashort cfDNA fragments confounding the analysis of whole plasma sequencing data for individual CTCF binding site occupation. ChIP-seq analysis removes this interference and facilitates detailed occupancy analysis.

ChIP-seq analysis of a plasma nucleoprotein by ChIP is dependent on a number of critical factors, including the amount and stability of nucleoprotein available in plasma for analysis and the sensitivity of DNA library preparation methods used for ultrashort cfDNA fragments. Plasma cell free ChIP-seq of cell free nucleosomes provides low numbers of ChIP-seq reads, for example the 4.5 million average reads reported for ChIP-seq of plasma nucleosomes carrying the histone modification H3K4Me3 [52]. The number of aligned ChIP-seq reads observed here for plasma cfCTCF-DNA of 877,664 was lower still. This likely relates to the 40,000-60,000 CTCF protected loci in the genome compared to the approximately 20 million nucleosomes present in each human cell of which 1-2% are modified by H3K4 trimethylation, as well as the greater stability of nucleosomes compared to low affinity cfCTCF-DNA nucleoproteins. Critical elements to enable ChIP-seq analysis of cfCTCF-DNA were formaldehyde crosslinking of plasma to protect cfCTCF-DNA from the destabilizing effects of EDTA zinc ion depletion on ZFP DBDs, and the use of single strand DNA library preparation to maximize sensitivity for ultrashort sub-100 bp CTCF associated cfDNA fragments.

Histone octamers bind to DNA with no sequence specificity to form near meta-stable nucleosome complexes. In contrast, TFs bind to DNA in a sequence dependent manner with highly variable affinities and dissociation rates to form TF-DNA complexes with residence times varying from seconds, for dynamic TFs, to hours for stable TFs [22, 27, 53]. CTCF has at least tens of thousands of binding sites in the human genome bearing a wide range of motif sequences with varying DNA binding affinities according to their concordance with the consensus DNA-binding motif. High affinity motifs bear sequences close to the consensus sequence and are reported to be associated with topologically associated domain loops (TADs) and the control of chromatin architecture [54]. CTCF-DNA nucleoproteins including high affinity motif sequences have low dissociation rates with long residence times of up to 22 hours but are a minority of binding sites [21, 22]. Lower affinity motifs are less similar to the consensus sequence and are associated with transcriptional regulation. Lower affinity CTCF-DNA nucleoproteins have higher dissociation rates, shorter residence times and make up a majority of CTCF binding sites. Furthermore, CTCF binding affinity for a motif is not fixed but is reduced by methylation of cytosine bases within the CTCF motif and modulated by binding with transcription co-factors as well as by interaction with RNA molecules [55]. CTCF binding to cohesin is associated with TAD formation and insulator functions and stabilizes CTCF-DNA binding leading to increased residence times [3, 55].

CTCF binding is dysregulated in cancer and plays key roles in oncogenesis and tumor development through a variety of mechanisms [1–13]. Gain of CTCF binding may be induced by oncogenic transcription factors such as NOTCH1 in acute lymphoblastic leukemia [6]. In breast cancer cell lines CTCF binding may influence estrogen receptor (ER) binding to chromatin thereby modulating ER mediated altered gene expression [7]. CTCF Hi-C studies reported more than 500 cancer specific CTCF loci in triple-negative breast cancer cells that were undetectable, or barely detectable, in normal tissues [8]. CTCF is also involved in dysregulation of MYC oncogene expression in a variety of cancers including in prostate cancer, where MYC binds with CTCF at MYC/CTCF common binding regions, resulting in stabilization of CTCF-DNA at these regions and three-dimensional re-organization of chromatin [56, 57]. However, high affinity plasma CTCF ChIP-seq peaks observed here did not overlap published cancer specific sites.

Similarly, ChIP-seq studies of fragmented cellular chromatin report both CTCF sites that are cell specific in terms of occupation as well as others that are ubiquitously occupied in many cell types [1, 2, 17–20]. The majority of plasma ChIP-seq peaks identified here overlapped with peaks observed in many cell types in ENCODE data sets. Moreover, the number of patient samples studied was too small to investigate cell specific CTCF binding patterns. However, future studies involving larger patient cohorts may establish plasma CTCF binding patterns that are characteristic of the cell of origin of cfCTCF-DNA.

We characterized a plasma cfDNA fragment as CTCF bound if it was bound by an anti-CTCF antibody, was sub-nucleosomal in size and included an approximately centrally located CTCF binding motif sequence with upstream and downstream flanking sequences that identified a mappable CTCF binding locus in the human genome. Using a first developmental cohort of patient samples, we selected an example panel of 14 high affinity CTCF loci sequences at a false discovery rate adjusted p ≤ 0.01 that, in combination, gave rise to at least one ChIP-seq peak in all 49 cancer plasma samples tested, but no peaks for 21 samples from healthy subjects or those with an inflammatory condition. Testing of the panel for the 14 ChIP-seq peaks in a subsequent, blinded validation sample cohort produced a negative result for all healthy samples tested, and a positive result in 13 of 14 samples from stage III/IV cancer patients. We here tested a minimal panel for liquid biopsy whose 14 members were selected purely empirically as giving a positive result in the samples tested in the development cohort with no false negative results. However, we subsequently observed that CTCF ChIP-seq peaks for thousands of high affinity CTCF binding loci represent tumor derived ctDNA and are not present for healthy samples. The number of possible panels is therefore very large and analysis of any panel selected by whole genome sequencing or targeted gene sequencing, or analysis of all loci, would be possible.

Whilst the small sample cohort sizes tested to date are insufficient to make a determination of sensitivity and specificity for cancer, it is clear that high affinity cfCTCF-DNA nucleoproteins may be useful novel biomarkers in oncology and hold promise for early cancer detection using either whole genome sequencing or targeted gene sequencing for panels of CTCF sequences in plasma ChIP isolates. Analysis of ctDNA bound at high affinity to CTCF in plasma has one great advantage for liquid biopsy in that a combination of chemical, in vitro, and bioinformatic, in silico, processes allow the removal of all background non-tumor cfDNA from analysis.

cfCTCF-DNA is not the only TF-DNA nucleoprotein that occurs in the circulation [14–16]. Indeed, the occupancy of more than 370 cancer cell derived TFs, known to have more than 10,000 BS in the human genome, have been inferred by nucleosome positioning studies of plasma cfDNA and a number of these have been directly footprinted in plasma [58, 59]. It is likely that at least some of these TFs are represented by high affinity cfTF-DNA nucleoproteins that may similarly be amenable to analysis by plasma ChIP-seq methods. For example, liquid biopsy analysis of androgen receptor or estrogen receptor binding to their respective response elements are likely to be informative in prostate and breast cancer respectively [60–62].

In conclusion, we report the first isolation and analysis of cfCTCF-DNA from plasma. We observed that cfCTCF-DNA in plasma samples from healthy subjects and cancer samples differs both qualitatively and quantitatively. Plasma CTCF ChIP-seq has enabled the first liquid biopsy analysis of CTCF binding site occupancy and identified a novel class of cfDNA biomarkers comprising ultrashort plasma cfCTCF-DNA loci sequences. The cfCTCF-DNA results obtained here identified 49 of 49 cancers tested in a developmental cohort, and 13 of 14 cancers in a validation sample cohort, with no false positive results. Although the sample numbers tested in this first proof-of-concept experiment were small and need to be confirmed in studies of larger independent clinical cohorts, the results show feasibility for a novel liquid biopsy approach involving analysis of ultrashort cfDNA fragments bound to CTCF at high affinity. These ultrashort plasma cfDNA fragments may be detected by PCR or targeted gene sequencing of plasma cfCTCF-DNA isolated by ChIP or by whole genome sequencing.

Limitations of the study

This is a first description of the isolation of plasma CTCF-cfDNA nucleoproteins. We characterize these nucleoproteins and demonstrate proof-of-concept for their potential use as a new class of cfDNA biomarkers. The patient sample results reported provide proof-of-concept for the use of plasma CTCF-cfDNA nucleoproteins as biomarkers, However, the patient cohorts studied to date were small, not age and sex matched and were of insufficient size to make a determination of sensitivity and specificity. It is also unlikely that the sequence panels selected based on these small cohorts were optimal. Further studies to refine biomarker panels and any potential clinical utility of the biomarkers will require further investigations in larger numbers of patients. We have not assessed any potential patient selection, monitoring, or prognostic associations.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Conceptualization: J.M.; Methodology: J.M., D.P. and B.C.; Software and Bioinformatics: T.B., B.P.B., and J.-V.T.; Investigation: D.P., B.C., A.G. and M.C.; Formal Analysis: J.M., D.P., B.P.B. and T.B.; Visualization: J.M.; Writing—Original Draft: J.M.; Writing—Review & Editing: A.R., D.P., M.H. and T.B.; Data Curation: D.P.; Funding Acquisition: J.M., M.H. and A.R.; Supervision: J.M., D.P. and M.H. All authors reviewed the manuscript.

Funding

This work was funded in part by the Walloon Region (convention #8515). Funding was provided by Volition (London, UK).

Data availability

The sequencing data supporting this study have been deposited in the European Genome-phenome Archive (EGA) under study accession EGAS50000001937.

Declarations

Conflict of interest

With the exception of J.-V.T., B.P.B., and A.G., all authors are current employees of Volition. B.P.B. was a consultant with Volition at the time the work was performed. M.C., B.P.B., M.H., A.R., T.B. and J.M. are shareholders of Volition. A patent application has been filed for aspects of the methods disclosed here.

Ethical approval

Blood samples obtained from patients and healthy donors were sourced from certified biobanks, BioIVT and Indivumed, conforming to the Declaration of Helsinki and the Belmont Report with written informed consent. Patient samples purchased from BioIVT LLC were collected under approval from the IRB board at WCG relating to the study titled Prospective Collection of Biological Specimens from Subjects Presenting at Specimen Donation Centers for Research, sponsored by Seratrials LLC. Patient samples purchased from Indivumed GmbH were collected under approval from the IRB board Ethik-Kommission der Arztekammer Hamburg relating to the study titled Indivumed Biospecimen Bank, sponsored by Indivumed Inc.

Consent for publication

Not applicable.