Plasma cell-free DNA chromatin immunoprecipitation profiling depicts phenotypic and clinical heterogeneity in advanced prostate cancer

Abstract

Cell phenotype underlies prostate cancer presentation and treatment resistance and can be regulated by epigenomic features. However, the osteotropic tendency of prostate cancer limits access to metastatic tissue, meaning most prior insights into prostate cancer chromatin biology are from preclinical models that do not fully represent disease complexity. Noninvasive chromatin immunoprecipitation of histones in plasma cell-free in humans may enable capture of disparate prostate cancer phenotypes. Here, we analyzed activating promoter- and enhancer-associated H3K4me2 from cfDNA in metastatic prostate cancer enriched for divergent patterns of metastasis and diverse clinical presentation. H3K4me2 density across prostate cancer genes, accessible chromatin, and lineage-defining transcription factor binding sites correlated strongly with circulating tumor DNA (ctDNA) fraction—demonstrating capture of prostate cancer-specific biology and informing the development of a statistical framework to adjust for ctDNA fraction. Chromatin hallmarks mirrored synchronously measured clinico-genomic features: bone versus liver-predominant disease, serum PSA, biopsy-confirmed histopathological subtype, and RB1 deletions convergently indicated phenotype segregation along an axis of differential androgen receptor activity and neuroendocrine identity. Detection of lineage switching after sequential progression on systemic therapy in select patients indicates potential utility for individualized resistance monitoring. Epigenomic footprints of metastasis-induced normal tissue destruction were evident in bulk cfDNA from two patients. Finally, a public epigenomic resource was generated using a distinct chromatin marker that has not been widely investigated in prostate cancer. These results provide insight into the adaptive molecular landscape of aggressive prostate cancer and endorse plasma cfDNA chromatin profiling as a biomarker source and biological discovery tool.

Introduction

Understanding cancer chromatin hallmarks is important for guiding drug development and biomarker discovery. For example, histone methylation profiling can nominate new drug targets and enables pharmacodynamic monitoring of epigenome-targeted therapies in development (e.g., EZH2 and histone deacetylase inhibitors) (1). Epigenomic features may also modulate treatment response and can mediate adaptive acquired treatment resistance to standard-of-care and emerging new therapies. In prostate cancer, epigenomic dysregulation is thought to influence tumor cell biology across multiple disease settings. The androgen receptor (AR) binding landscape is reprogrammed during initial tumorigenesis and later transition to lethal metastatic castration-resistant prostate cancer (mCRPC) (2,3). In mCRPC, epigenomic remodeling can facilitate treatment resistance via lineage plasticity and emergence of AR-independent neuroendocrine disease (4–9).

However, nearly all insight into prostate cancer chromatin regulation is derived from preclinical model systems (2,10–16). Limited access to relevant metastatic tissue has historically constrained characterization of epigenomic disease features, forcing reliance on in vitro model systems that poorly recapitulate the complexity or biological spectrum of clinical disease. In prostate cancer, challenges with obtaining metastatic tissue are intensified as bone is the most common site of metastasis, which is technically difficult to biopsy (~50% failure rate (17)) and typically involves preanalytical decalcification that destroys cells. These practical complications are epitomized by the complete absence of chromatin immunoprecipitation data from bone metastases across any solid human cancer. Consequently, epigenomic correlates of human mCRPC metastatic organotropism and clinical phenotype remain largely unexplored.

Plasma cell-free DNA (cfDNA) from dying cells harbors multiple cell-of-origin epigenomic hallmarks. This includes direct posttranslational DNA modifications (e.g., 5(hydroxy)methylcytosine (4–6,8,9)) and indirect proxies of chromatin architecture (e.g., fragmentomic footprints of transcription factor activity and nucleosome positioning (18–20)). Intriguingly, most cfDNA remains bound to histones—which was recently demonstrated to be analyzable via chromatin immunoprecipitation followed by sequencing (cfChIP-seq)—unlocking an additional reservoir of epigenomic information previously only accessible via invasive tissue profiling (21–25). Importantly, DNA alterations (in ctDNA or tissue) incompletely inform tumor cell characteristics, adaptive treatment-resistance mechanisms, and clinical phenotype, underscoring a need for practical in vivo phenotyping techniques that capture a broader spectrum of clinically-relevant tumor characteristics. Plasma cfChIP-seq may circumvent limitations of chromatin profiling in both tissue and model systems, but has not been investigated in clinically-annotated prostate cancer patients enriched for disparate radiographic disease features, nor with methods that assess the potential confounder of variable ctDNA fraction (ctDNA%).

In this hypothesis-generating study, we explore the potential for plasma cfChIP-seq to inform on the biological and clinical heterogeneity of mCRPC. Identification of clear biological differences between rational dichotomous patient subgroups in our cohort underscores the capacity of dual ctDNA genomic and epigenomic profiling to reveal unique insight into metastatic biology. We also generate a new public plasma cfChIP-seq resource using a distinct chromatin marker (H3K4me2) that has not been widely applied to metastatic prostate cancer. Our results suggest that plasma cfChIP-seq is poised to accelerate the discovery of clinically-relevant biological states in mCRPC.

Materials and Methods

Cohort

We profiled patients from a prospective province-wide plasma cfDNA biobanking program at the Vancouver Prostate Centre and BC Cancer. All patients had histologically-confirmed prostatic adenocarcinoma (high-grade neuroendocrine and/or small-cell components permitted) or metastatic urothelial cancer and radiographic evidence of metastatic disease by conventional imaging (CT or bone scintigraphy). Plasma cell-free DNA (cfDNA) and matched white blood cell (WBC) DNA samples were collected prior to initiation of systemic treatment (any line of therapy) and at subsequent clinical progressions in select patients. cfDNA and WBC samples were subjected to deep targeted cfDNA sequencing using an established panel of 77 prostate cancer relevant genes. To enhance tumor-specificity of cfChIP-seq, we selected samples for plasma cfChIP-seq based on harboring high overall cfDNA tumor purity from initial targeted sequencing (i.e. approximately >20% ctDNA% where feasible). As additional controls, we performed plasma cfChIP-seq on 11 ctDNA-negative samples from prostate cancer patients (provincial biobank) and 8 plasma and 2 buffy coat samples from healthy donors (NIH). The 11 control ctDNA-negative samples were collected from mCRPC patients but deep targeted sequencing did not detect evidence of ctDNA (at above 0.5%) based on somatic allele frequencies and genome-wide copy number profiles.

For patients with prostate cancer, clinical data was retrieved from retrospective chart review and included patient demographics, clinical, pathological and laboratory features at the time of prostate cancer diagnosis and time-matched to each cfDNA collection, as well as time-to-event outcomes (Supplementary Table 1). We collected bone scintigraphy and CT scan data taken closest to time of cfDNA collection, with preference for scans preceding cfDNA collection by at most 30 days, where feasible. Images were independently reviewed by a nuclear medicine physician (I.B.), blinded from plasma cfChIP-seq data and patient outcomes, using standardized predefined criteria addressing anatomic location of metastases, sizing and enumeration of lesions, and lesion morphology (e.g., bone metastases with associated soft tissue component). Imaging features were selected to capture established radiographic hallmarks of neuroendocrine/small-cell prostate cancer (i.e. bulky (≥5 cm) lymphadenopathy, lytic bone lesions, visceral predominant disease)(26) and enable patient dichotomization by disease burden using CHAARTED criteria (27). Sex and/or gender are not relevant for any findings within this study and were therefore not incorporated into study design, clinical data collection, nor execution of analyses. All samples are de-identified at time of collection, and all researchers are blind to patient gender identity and gender presentation.

Approval for collection and profiling of patient samples was granted by the University of British Columbia Research Ethics Board (certificate numbers H18-00944, H16-00934) and NIH Institutional Review Board (99-CC-0168), and all samples were de-identified prior to analysis. The study was conducted in accordance with the Declaration of Helsinki, and written informed consent was obtained from all patients prior to enrollment.

Genomic and cell-free DNA isolation

Blood samples were collected using Streck Cell-Free DNA BCT^® tubes (2–3× 9mL blood tubes collected per patient) and processed at room temperature at the Vancouver Prostate Centre, adhering to the manufacturer’s guidelines. To ensure proper storage, resulting plasma samples were divided and stored at −80°C in labeled conical tubes. Similarly, buffy coat samples were placed in a single labeled conical tube and stored at −80°C. For DNA extraction, the Promega Maxwell RSC Blood DNA Kit and QIAamp Circulating Nucleic Acid Kit were utilized to extract white blood cell DNA (WBC DNA) and cell-free DNA (cfDNA), respectively. A buffy coat input of 30μL was used for WBC DNA samples, which were then eluted in 50μL of Promega Elution Buffer (REF-A828D). For total cfDNA extraction, 6mL of plasma was utilized following the manufacturer’s protocol, and the elution was done in 60μL of double-distilled water. To quantify the concentrations of WBC DNA and cfDNA, the QuantiFluor ONE dsDNA kit and Quantus Fluorometer from Promega was employed. However, in cases where the total yield of cfDNA exceeded 50ng/mL of plasma, further assessment for WBC DNA carry-over was conducted using a 1.3% SYBR-Safe agarose gel.

Target capture and sequencing

For WBC DNA samples, library preparation used 50ng inputs. WBC DNA was sheared to a median fragment length of 180bp through enzymatic fragmentation, followed by end-repair, A-tailing, and overnight adaptor ligation, incorporating 3-bp IDT xGen CS Unique Molecular Identifiers (UMI). cfDNA sample inputs ranged from 10 to 100ng, depending on the overall yield from cfDNA extraction. Fragmentation was not performed on cfDNA. For both WBC DNA and cfDNA, following adaptor ligation PCR amplification was carried out using the KAPA HyperPrep Kit (Roche) with IDT Unique Dual Index (UDI) primers, performing 5 to 8 cycles. Library quantification was conducted with the NanoDrop spectrophotometer, and each library underwent quality control by running it on a 1.3% SYBR-Safe agarose gel. To create pools, the libraries were multiplexed, resulting in single pools with a combined mass of 2.5μg. These library pools were then hybridized to a custom-designed KAPA HyperChoice probe capture panel for a minimum of 16 hours at 55°C. The panel was designed to capture coding regions of 77 prostate cancer-relevant genes, as well as introns and flanking regions of selected clinically-relevant genes such as AR, TP53, PTEN, and RB1. Additionally, our panel included regularly spaced probes distributed across the genome that capture heterozygous germline SNPs of common population frequency. The subsequent steps of wash, recovery, and amplification of the captured regions were performed following the KAPA HyperCap Workflow protocols. Final libraries were purified using KAPA HyperPure Beads and quantified using the QuantiFluor ONE dsDNA kit and Quantus Fluorometer (Promega). The pools were then diluted to 4nM and subjected to sequencing on Illumina machines.

Plasma cell-free nucleosome chromatin immunoprecipitation

Antibodies were coupled to M-270 epoxy dynabeads (Thermofisher) following manufacturer’s instructions at a concentration of 5 ug of antibody per mg of beads. Antibodies used: H3K4me2 (Diagenode; catelogue #: C15410035), H3K4me3 (Diagenode; catelogue #: C15410003, RRID: 2924768), H3K27me3 (Diagenode; catelogue #: C15410195, RRID: AB_2753161), H3K36me3 (Diagenode; catelogue #: C15410192, RRID: AB_2744515) (Supplementary Tables 2,3). All antibodies used in this study have experimentally-validated low cross-reactivity to different histone posttranslational modifications, as evidenced by dot blot assays provided by the vendor plus systematic independent third-party validation (of H3K4 methylform specificity only) (28). This specific H3K4me2 antibody has also been utilized across multiple prior in vitro and human studies (21,29). Successful coupling was determined by immunoblotting of supernatant before and after coupling. Antibody coupled beads were added to 300uL to 2mL plasma based on concentration of cfDNA and incubated overnight at 4°C. A fraction of the plasma was collected as input. The beads were washed three times with buffer A (50 mM Tris pH 7.5, 2mM EDTA pH 8.0, 150mM NaCl, 0.1% sodium deoxycholate, 1% Triton X-100), three times with buffer B (50 mM HEPES pH 7.6, 1mM EDTA pH 8.0, 500mM LiCl, 0.7% sodium deoxycholate, 1% NP40), and twice with TE (10mM Tris pH 8.0, 1mM EDTA pH 8.0). DNA was purified following proteinase K and RNase treatment using the Monarch DNA Cleanup kit (NEB). Libraries were prepared using the NEBNext Ultra II Library Preparation kit and sequenced on the Illumina NextSeq2000 platform at the CCR Genomics Core at the National Cancer Institute, NIH, Bethesda MD.

Buffy coat fraction was washed with phosphate buffered saline and lysed with lysis buffer (10mM Tris pH 8.0, 1mM EDTA, 0.5% SDS). DNA was sheared using Bioruptor Pico and clarified supernatant diluted with dilution buffer (20mM Tris pH 8.0, 2mM EDTA, 150mM NaCl, 1% Triton) and a fraction was taken for input DNA. Chromatin was incubated with indicated antibodies overnight at 4°C. Protein A/G Dynabeads (Thermofisher) were added and incubated for 1 hour at 4°C. Beads were washed 6 times with wash buffer (50mM HEPES pH 7.6, 1mM EDTA, 500mM LiCl, 0.7% sodium deoxycholate, 1% NP40) followed by elution with 100mM sodium bicarbonate, 1% SDS. DNA was purified following proteinase K and RNase treatment using the Monarch DNA Cleanup kit (NEB). Libraries were prepared using NEBNext Ultra II Library Preparation kit and sequenced on the Illumina NextSeq2000 platform at the CCR Genomics Core at the National Cancer Institute, NIH, Bethesda MD.

Chromatin immunoprecipitation in LNCaP cell line

LNCaP cell line (RRID:CVCL_1379) was purchased from ATCC and grown in RPMI (Gibco) supplemented with 10% fetal calf serum (Gibco), 100U/mL penicillin, and 100ug/mL streptomycin (Gibco). Cell lines were authenticated by short tandem repeat profiling using Promega GenePrint 10 (Laragen) and were tested for mycoplasma every 6 months using MycoAlert Mycoplasma detection kit (Lonza)—cell lines were last tested in January 2023 and were negative. Cells were thawed and passaged for up to 6 months for experiments before thawing new frozen vial. Ten million cells were fixed using 1% formaldehyde (Thermofisher) for 10 minutes at 37°C and quenched with glycine. Chromatin was sheared in lysis buffer (10mM Tris pH 8.0, 1mM EDTA, 0.5% SDS) using the Bioruptor Pico to 300–500 base pairs. Immunoprecipitation was done overnight with 2ug of indicated antibodies. Protein A/G Dynabeads (Thermofisher) were added and incubated for 1 hour at 4°C. Beads were washed 6 times with wash buffer (50mM HEPES pH 7.6, 1mM EDTA, 500mM LiCl, 0.7% sodium deoxycholate, 1% NP40) followed by elution with 100mM sodium bicarbonate, 1% SDS. DNA was purified following proteinase K and RNase treatment using the Monarch DNA Cleanup kit (NEB). Libraries were prepared using NEBNext Ultra II Library Preparation kit and sequenced on the Illumina NextSeq2000 platform at the CCR Genomics Core at the National Cancer Institute, NIH, Bethesda MD.

Sequence alignment

Illumina adapters were trimmed using cutadapt (30) with the following arguments: --interleaved -m 20 -a AGATCGGAAGAGC -A AGATCGGAAGAGC --overlap 1. Low-quality bases from the 3’ end were trimmed using fasta trim by quality from seqkit. Bases with a quality score less than 20 were masked. Trimmed and masked FASTQ files were then aligned to the human genome (hg38) using Bowtie2 with the following settings: -X 1000 --score-min L,0,−0.6 --ignore-quals. Bam files were deduplicated using samblaster and sorted using samtools.

Somatic and germline variant calling

Our somatic and germline variant calling methodologies have been previously described (18). Briefly, for somatic variant calling we used Mutato with arguments --alt-reads=8 --alt-frac=0.05 and modified the resulting matrix of candidate mutations using in-house scripts. We retained mutations with variant allele frequency (VAF) ≥20× the per-position background error rate (empirically derived from a large pool of ctDNA-negative cfDNA samples) and ≥3× the VAF in the patient-matched WBC sample. For germline variant calling, we retained WBC variants with a VAF of ≥0.15, along with ≥5 alternate reads. All candidate germline mutations were required to have a VAF of ≥20× the background error rate at the respective genomic position. All candidate variants were manually curated using Integrative Genomics Viewer (IGV). All somatic mutations are reported in Supplementary Table 4.

ctDNA fraction and copy number estimation

Copy number status and genome equivalent ctDNA fractions (i.e. the typical measure of tumor purity, defined as the fraction of plasma cfDNA genome equivalents derived from cancer cells, which is equivalent to the fraction of ctDNA fragments in regions with normal ploidy and is reported by standard tools such as ichorCNA (31)) were generated using the input DNA controls from the chromatin immunoprecipitation protocol (median 0.96× lpWGS) together with genome-wide heterozygous SNP allele fractions from deep targeted sequencing, leveraging previously validated methodology (see Supplementary Methods of Herberts et al.) (18).

Tumor- and normal-derived cfDNA are associated with distinct epigenomic properties, and both contribute to cfChIP-seq. At a genomic region $R$, the observed cfChIP-seq signal $S_{RO}$ is composed by

$$S_{RO}=F_R{S}_{RC}+\left(1-F_R\right)S_{RN}$$

where $F_R$ is the fraction of ctDNA fragments from $R$ in the input cfDNA sample, and $S_{RC}$ and $S_{RN}$ are the pure signals of cancer and normal cells in $R$, respectively. Comparing epigenomic features of individual regions within or between samples therefore necessitates careful adjustment for local ctDNA fraction. Correction using sample ctDNA genome equivalents (i.e. an overall estimate of ctDNA fraction) is potentially inaccurate since somatic copy alterations additionally impact the ratio of tumor to normal cfDNA at specific loci. We leveraged our whole-genome copy number information to calculate local ctDNA fraction $F_R$ for each non-overlapping 1MB genomic region $R$:

$$F_R=F_G{C}_{RC}/\left(F_G{C}_{RC}+\left(1-F_G\right)C_{RN}\right)$$

where $F_G$ is the genome equivalent ctDNA fraction of the sample, and $C_{RC}$ and $C_{RN}$ are the copy numbers of $R$ in cancer and normal cells, respectively (2 for autosomes and 1 for allosomes in normal cells); $F_G$ and $C_{RC}$ are empirically inferred from whole-genome copy number fitting. Within the figure panels of this study, we used either local ctDNA% (i.e. specific to individual genomic features of interest) or genome equivalent ctDNA% (i.e. as a measure of overall sample tumor purity) as appropriate—this distinction is made clear in the figure captions. Per gene copy number estimates are provided in the Supplementary Table 5.

Plasma cfChIP-seq data normalization

We first counted 50–500bp fragments in non-overlapping 100bp genomic windows (spanning the entire genome) for all samples, and used these 100bp resolution fragment counts downstream in all analyses except for motif enrichment analyses. We then applied a standard GC correction to all input control samples, by modeling the relationship between GC fraction and fragment counts using a fourth-degree polynomial, and dividing the fragment count of each window by the polynomial-fit value.

For TSS and gene body analyses, we used 17,774 genes defined in MANE GRCh38 v0.93. For transcription factor (TF) binding sites (TFBS), we used narrowPeak files of 594 ENCODE TF ChIP-seq experiments for 457 unique TFs (32) done in prostate, blood or breast tissues/cell-lines (Supplementary Table 6), and an AR ChIP-seq experiment done on primary prostate cancer specimens (Supplementary Table 7) (33). Gene bodies with <1 average fragments between samples were omitted (n=8071 valid genes) to increase the biological signal to background noise ratio. TSS and TFBS fragments were counted in a 2kB neighborhood around the TSS (n=17,486 TSS neighborhoods had >0 total fragments). For each publicly-available TFBS peak set, we filtered out peaks containing 0 fragments across all our cfChIP-seq samples, and for each sample calculated the average count between the remaining peaks.

After initial preprocessing to remove uninformative regions, we divided cfChIP-seq fragment counts for each sample by the same sample’s input DNA lpWGS fragment counts to normalize for copy number alterations and other sources of local coverage variation. Matched lpWGS input controls were sequenced to a relatively low depth (median 0.96× read depth), resulting in a high degree of relative coverage differences that is magnified when surveying small genomic intervals. Naively dividing cfChIP-seq fragment counts by matched size lpWGS input regions may diminish signal-to-noise ratio in smaller regions (e.g. individual TSSs). To mitigate this risk, we tailored our normalization strategy to each specific analysis. For analyses focused on measuring histone marker counts within individual TSS regions, open chromatin regions (narrowly defined intervals derived from prior ATAC-seq data), or within a set of broadly called consensus peaks (described below), each cfChIP-seq count was divided by the average lpWGS input control count within a 1MB neighborhood. For analyses focused on measuring histone marker counts across gene bodies and TFBS, cfChIP-seq fragment counts were normalized using coordinate-matched regions in matched input lpWGS samples. Since some TFs had multiple available public ChIP-seq peak sets, we averaged the fragment counts corresponding to different sets of binding sites for the same TF.

Next, we normalized samples for sequencing depth, which is necessary for comparing samples or sample groups. Scaling samples by total sequencing depth alone (default behavior of DiffBind (34)) left large differences between samples at all regions of interest (Supplementary Figure 1). Since total sequencing depth consists of both true biological signal and background noise, any differences could be due to genome-wide changes in histone modification (biological), or differences in background noise (technical, e.g. due to non-specific immunoprecipitation). Given that background noise is known to vary between ChIP-seq experiments, we reasoned that it is the most probable reason for the global differences between samples. Therefore, we applied a similar strategy to DiffBind “RLE” option (default normalization by DeSeq2 (35), scaling the samples based on a set of consensus control regions instead of the whole genome. For gene bodies, TSSs, and TFBSs, the control regions were the aforementioned 8071 genes, 17,486 genes, and 458 TFs, respectively. For consensus peaks (solely leveraged for genome-wide principal component analysis) and open chromatin regions, we selected 2000 of the 100bp windows as control peaks through the following iterative procedure (Supplementary Figure 2A):

1. Calculating the minimum fragment count across all samples for each 100bp window.

2. Selecting 20,000 consensus peaks that had the highest minimum fragment count (across all samples, determined by step #1)

3. Selecting 2000 of these 20,000 consensus peaks harboring the least between-sample logarithmic variance

Then, for each sample, we scaled all regions so that the mean of these 2000 peaks was 1.0, and repeated the peak calling and scaling a total of 10 times so that the final control peak set was minimally affected by sequencing depth differences between samples.

To carry out GC bias correction for the plasma ChIP-seq samples, we observed that the standard GC correction strategy we used for input DNA samples was inadequate, because it assumes the underlying true fragment depth to be equal in all genomic regions. In ChIP-seq data, the expected depth depends mainly on immunoprecipitation, which can lead to depth differences in 1000-folds. Consensus control regions, however, have similar biological signal between samples, and thus their between-sample ratios are expected to be 1. Therefore, instead of scaling samples by just the median of ratios (like DiffBind “RLE”), we scaled them by a GC curve fit to the ratios to simultaneously correct between-sample GC bias differences. Since normalization with 2000 control peaks was used also for within-sample comparisons (cancer type specific open chromatin signatures), we applied GC correction to the raw counts instead of ratios. Relationships between GC and fragment counts were very typical (Supplementary Figure 1B,C). Our stepwise procedure for scaling the samples is:

1. Calculating log ratios for each sample against the mean of all samples (Supplementary Figure 2B) (except for the 100bp wide control peaks: Supplementary Figure 2A)

2. Modeling the relationship between the log ratios and GC fractions (using a fourth-degree polynomial fit, as above with input controls)

3. Reducing the fragment count of each region with GC-fraction-matched model count (Supplementary Figure 2B) (dividing instead of reducing in the case of 100bp wide control peaks: Supplementary Figure 2A).

Log ratios were converted back to regular counts after normalization (Supplementary Figure 2B). Normalized counts and local ctDNA fractions of each TF, TSS and gene body are in Supplementary Tables 8–10, respectively.

Interaction testing for differential histone modification status

Plasma cfDNA samples from cancer patients are combinations of non-tumor cfDNA (usually originating from hematopoietic cells) and ctDNA from cancer cells. Therefore, the ratio of the two contributors to bulk plasma cfDNA is a major factor affecting measured histone modification signals. This factor is particularly important to consider when comparing groups of samples (e.g. from prostate versus bladder cancer) so that differences in signal between groups can be reasonably attributed to differences in histone modifications in the cells contributing to ctDNA, rather than due to differences in the local or global ctDNA% between groups. If, for example, cancers of samples A and B have equal histone modification status at a feature of interest, we can observe a relative cfChIP-seq signal difference up to their relative ctDNA% difference (e.g. 2-fold difference in ctDNA%, up to 2-fold difference in observed cfChIP-seq signal).

To measure differential modification between normal cfDNA and ctDNA, we fit a linear model to fragment counts (in select regions, e.g. gene bodies, TSS-proximal intervals, etc.) of cfDNA samples:

$$Y={\beta }_0+{\beta }_{1}F$$

where $Y$ is the fragment count of the sample, and $F$ is the local ctDNA fraction (Supplementary Figure 2C1). The inferred “pure” normal cfDNA and ctDNA counts (i.e. scenarios where the sample is 100% normal cfDNA or 100% ctDNA) are:

Normal cfDNA: ${\beta }_0$ (i.e. the y-intercept of the linear model fit)

ctDNA: ${\beta }_0+{\beta }_1$ (i.e. where the linear model fit line transects the F=1 (100% ctDNA) boundary condition)

To visualize the difference between normal cfDNA and ctDNA of each TSS/gene body/TF, we plotted the log ratios between pure ctDNA and pure normal cfDNA counts:

$$\mathrm{l}\mathrm{o}\mathrm{g}2\left(\left({\beta }_0+{\beta }_1\right)/{\beta }_0\right)$$

against Pearson correlations between ctDNA% and fragment counts (Supplementary Figure 2C2)

To test for differences between the cancers of two sample groups, we first selected only the highest ctDNA% sample from each patient (all ctDNA negative and healthy samples were included). We then calculated $\mathrm{F}$-test $\mathrm{p}$-values for linear models:

$$\text{H0:}Y={\beta }_0+{\beta }_{1}F$$

$$\text{H1:}Y={\beta }_0+{\beta }_{1}F+{\beta }_{2}FG$$

where $G$ is a categorical variable for the sample group (Supplementary Figure 2D1). The pure fragment counts of the two sample groups are:

$$\mathrm{G}\mathrm{r}\mathrm{o}\mathrm{u}\mathrm{p}1:{\beta }_0+{\beta }_1$$

$$\mathrm{G}\mathrm{r}\mathrm{o}\mathrm{u}\mathrm{p}2:{\beta }_0+{\beta }_1+{\beta }_2$$

When testing for a continuous variable instead of a categorical sample group (i.e. PSA), we swapped the categorical variable $G$ for a continuous one. We then visualized the differences by plotting sample group log ratios:

$$\mathrm{l}\mathrm{o}\mathrm{g}2\left(\left({\beta }_0+{\beta }_1+{\beta }_2\right)/\left({\beta }_0+{\beta }_1\right)\right)$$

against $\mathrm{F}$-test $\mathrm{p}$-values (Supplementary Figure 2D2), and further summarized the top features across multiple sample groupings with a dot matrix chart (Supplementary Figure 2D3).

H3K4me2 enrichment in cancer-specific open chromatin

Genomic regions with differential accessibility in clustered primary cancer types (identified through ATAC-seq) were obtained from a public TCGA dataset (36). Note that bladder urothelial carcinoma and prostate adenocarcinoma were reported in the primary publication as forming two distinct clusters in t-SNE space. We performed a bedtools intersect to extract the genomic intervals unique to each of the 18 clusters of 23 primary cancer types, followed by quantifying the normalized cfChIP-seq signal in these differentially accessible regions (Supplementary Figure 3A–C). To produce spatial distributions of normalized cfChIP-seq fragment counts, we counted fragments in neighborhoods around the midpoint of each differentially accessible genomic region, and averaged noncontiguous regions by matching positions with the same relative distance to the midpoint. In Supplementary Figure 3C (i.e. a heatmap where spatial information is collapsed), we calculated the average H3K4me2 intensity in a 1.5kb window surrounding the interval’s midpoint (i.e. interval [−750, +750]): average fragment counts within upstream and downstream flanking regions were subtracted from the midpoint values which spanned the 1000 bp region around the midpoint [−500, +500]. For the ctDNA-positive prostate and bladder cancer samples, H3K4me2 intensity was normalized relative to the cancer lineage-specific (i.e Supplementary Figure 3C columns) background distribution within ctDNA-negative, healthy control and WBC samples via subtraction.

Motif enrichment analysis

We performed peak calling on deduplicated un-normalized BAM files from cfChIP-seq experiments using MACS2 via the callpeak function with a p-value threshold of 0.01 and otherwise default parameters. We subsequently applied HOMER findMotifsGenome.pl on the narrowPeaks output file (from MACS2) to search for differentially enriched transcription factor binding motifs. We used parameters -size ‘given’ -len 8,10 to search for motifs across the entire coordinate space of called peaks. Known motifs identified by HOMER were then cross-referenced to only include those representing the top 100 most significantly enriched TFs in the four mCRPC epigenomic subtypes (CRPC-AR, -NEPC, -WNT, -SCL) identified from model systems in a published study (10). We then examined differences in subtype-specific TF motif enrichment across categorical sample groups split by clinical features, analyzing in aggregate all −log(p-value) enrichment scores for TFs assigned to the same subtype. Select comparisons were statistically quantified using the Fisher’s Exact Test with a p-value threshold of 0.05.

Statistical analyses and data visualization

Sample size was not predetermined for this retrospective exploratory study. The descriptive and hypothesis-generating nature of our study and lack of pre-specified analyses means that target sample sizes were not possible. When samples were excluded from sub-analyses, the rationale and denominators are clearly listed in the manuscript text and/or figure legend. We did not perform any analyses requiring patient/sample randomization.

Data analysis and statistical testing were conducted in Python 3.9 (using pandas 1.2.4, numpy 1.20.2, scipy 1.7.1, statsmodels 0.12.2), Julia 1.9.3 (GLM 1.8.3 and HypothesisTests 0.10.13), and R 4.0.3 (dplyr 1.0.7) languages. Visualizations were generated using matplotlib 3.3.4 (Python) and Julia (1.9.3). The following bioinformatics/genomic analysis software was used: cutadapt-1.11, seqkit-0.8, Bowtie-2.3.0, 2.3.4.3, samblaster-0.1.24, bedtools-2.25.0, samtools 1.12 (htslib 1.12), Picard 2.25.6, Mutato version 0.7, ANNOVAR (version 20191024), Picard 2.25.6, MACS2 (2.2.7.1), HOMER (4.11). The following databases were used for mutation annotation: COSMICv77, ExAC v0.3, Kaviar (2016-02-04 release), ClinVar (2019-03-06 release). All boxplots are centered at the median unless otherwise specified and display the interquartile range (IQR). Whiskers extend 1.5 × IQR past the quartiles; all individual datapoints are shown where feasible. All hypothesis tests were two tailed and required a 5% significance threshold.

Data availability

Human hg38 reference genome was downloaded from UCSC. Exon and TSS coordinates were obtained from RefSeq Matched Annotation from NCBI and EMBL-EBI (MANE). RNA-seq data of normal human tissues was obtained from Genotype-Tissue Expression (GTEx) (37). Tissue RNA-sequencing data from mCRPC metastatic biopsies and metastatic bladder cancer was obtained from previously published work (dbGaP study accession: phs001648.v2.p1 (38) and EGA accession EGAS00001004615 (39), respectively). Cancer-specific open chromatin regions were obtained from previously published work (36) leveraging TCGA primary cancer samples. AR TF binding sites were obtained from previous chromatin immunoprecipitation followed by sequencing of 13 primary prostate cancer tissue specimens (2,18,33). All other TF binding sites were downloaded from ENCODE (32).

De-identified plasma cfChIP-seq and targeted cfDNA and WBC DNA data generated in this study from patients with metastatic cancer are publicly available in the NCBI database (dbGaP) at phs003482.v2.p1. Sequencing data are available indefinitely for research use only under standard controlled access: data access inquiries should be directed to Dr. David Takeda (david.takeda@nih.gov) or Dr. Alexander Wyatt (alexander.wyatt@ubc.ca). Timeframe for data access will be subject to NCBI policy and process. All other raw data generated in this study are available in the article and/or Supplementary Data or upon request from the corresponding authors.

Code availability

Code available at https://github.com/sipolaj/plasma-cfChiP-seq-manuscript-code/. Custom tools to process sequencing data are available at https://github.com/annalam/seqkit/.

Results

cfDNA genomic and histone profiling in advanced prostate cancer

We interrogated 71 plasma and 7 buffy coat samples from 46 individuals including 34 patients with metastatic prostate cancer (mPCa; 32/34 mCRPC), 5 with metastatic bladder cancer, and 7 healthy controls. Patients were selected to represent diverse clinical phenotypes: including distinct organotropic patterns and burden of metastases, as well as adenocarcinoma versus biopsy-confirmed neuroendocrine features (NEPC) (Figure 1A; Table 1). Given our priority of exploring plasma cfChIP-seq as a research tool to understand metastatic cancer biology, all mPCa patients provided ≥1 high ctDNA fraction (ctDNA%) sample to enhance the tumor-specificity of epigenomic marks in bulk cfDNA (median ctDNA% among positives: 49.6%; interquartile range: 29.9–64.6%)—note that the ctDNA% in our cohort is considerably higher than that of unselected first-line mCRPC (median: ~1–5%) (40,41). 17 patients—including those with evidence for temporal clinical phenotype switch—provided serial cfDNA at sequential progressions on systemic therapy, and from 11 patients we had asynchronous ctDNA-negative cfDNA as controls (i.e. ctDNA% below ~1% as determined by somatic mutation analysis) (Supplementary Table 2).

Figure 1. Plasma cfChIP-seq in prostate cancer patients with comprehensive clinico-genomic annotation.

(A) Cohort characteristics (left) and multi-omic sequencing strategy (right). (B) Prostate cancer ctDNA genotypes plus select time-matched clinical characteristics and prior treatment exposure with documented resistance. (C) Per-sample (row) segmented copy number profiles of the AR gene and enhancer. Ligand binding domain (LBD) mutations and structural variants, and gene and enhancer copy number status are annotated. Blue vertical ticks represent individual per-sample rearrangement breakpoints; upper kernel density plot indicates aggregate breakpoint spatial density, expectedly converging on the AR gene body. (D) AR neighborhood histone modification density via multimodal plasma cfChIP-seq. Representative mCRPC (patient P13) and LNCaP (prostate cancer cell line) samples are enriched for activating markers relative to bladder cancer (patient P8) and a ctDNA-negative control sample (patient P6). Per-row vertical axis limit annotated (right). (E) Aggregate cfChIP-seq fragment count across all protein-coding gene bodies, incorporating all mCRPC cfDNA samples (n=58) and histone posttranslational markers analyzed. (F) Whole blood RNA expression correlates with TSS histone marker counts (±500bp neighborhood around the TSS) in a ctDNA-negative sample. Abbreviations: mod, moderate; mets, metastasis; ARPI, androgen receptor pathway inhibitor; LN, lymph node; ULN, upper limit of normal; TPM, transcripts per million; TES, transcriptional end site.

Table 1:

Prostate cancer cohort clinical characteristics (n=34)^a

Age at first plasma collection (years)	70 (65–72)
Stage at initial prostate cancer diagnosis
M0	12 (35)
M1a	4 (12)
M1b	14 (41)
M1c	4 (12)
ISUP Grade Group
1–3 (Gleason sum ≤ 7)	7 (21)
4–5 (Gleason sum ≥ 8)	24 (73)
Not assessable (e.g. metastatic biopsy, small cell histology)	2 (6)
No tissue biopsy at diagnosis	1 (3)
Number of serial plasma samples contributed
1	17 (50)
2	11 (32)
3	4 (12)
4	2 (6)
Disease state at first plasma collection
M0 CRPC	3 (9)
M1 CSPC	2 (6)
M1 CRPC	29 (85)
Lines of prior ARPI with resistance development at first plasma collection b
0	17 (50)
1	10 (29)
2	7 (21)
Lines of prior taxane with resistance development at first plasma collection b
0	28 (82)
1	6 (18)
2	0 (0)
Blood laboratory values at first plasma collection
PSA (ng/mL, serum)	37 (12–114)
Alkaline phosphatase / ULN	1.1 (0.8–2.4)
Lactate dehydrogenase / ULN	1.1 (0.8–1.9)
Hemoglobin (g/L)	107 (94–124)
Neutrophil-to-lymphocyte ratio	3.5 (2.0–5.4)
Disease burden and morphological characteristics by CT scan at first plasma collection (n=33) c
Bone metastases	24 (73)
Any lytic component	6 (25)
Any soft tissue component	5 (21)
Lymph node disease	19 (58)
By size
Non-bulky lymphadenopathy (<5cm)	16 (84)
Bulky lymphadenopathy (≥ 5cm)	3 (16)
By location
Pelvic only	3 (16)
Extrapelvic only	7 (37)
Both pelvic and extrapelvic	9 (47)
Visceral disease
Liver metastases	7 (21)
< 10 lesions	2 (29)
≥ 10 lesions	5 (71)
Lung metastases	5 (15)
< 10 lesions	3 (69)
≥ 10 lesions	2 (40)
Other (e.g. adrenal, pancreatic)	6 (18)
Disease burden by bone scan at first plasma collection (n=29) c
Bone metastases	21 (72)
By number
< 10 lesions	5 (24)
≥ 10 lesions	16 (76)
By location
Axial spine involvement only	4 (19)
Appendicular skeletal involvement only	1 (5)
Both axial spine and appendicular skeletal involvement	16 (76)

^aData are median (IQR), or n (%); percentages reflect proportion of patients with complete data for the given variable.

^bResistance to ARPI and taxanes as defined by PSA, radiographic, or PSA progression whilst receiving therapy.

^cNot all patients underwent both conventional CT and bone scan prior to first plasma collection. Therefore, disease burden is separately reported by imaging modality, with percentages reflecting denominators that had complete data.

Abbreviations: ADT, androgen deprivation therapy; ARPI, androgen receptor pathway inhibitor; CSPC, castrate-sensitive prostate cancer; CRPC, castration-resistant prostate cancer; ISUP, International Society of Urological Pathology; PSA, prostate-specific antigen; ULN, upper limit of normal.

For mPCa patients, plasma cfDNA collections were time-matched to radiographic and clinical evaluation of disease prior to initiating subsequent therapy. Manual assessment of bone scintigraphy and computed tomography (CT) imaging enabled objective quantification of metastatic disease: 61% (19/31; n=3 without bone scan data) of patients had >10 bone lesions, 39% (13/33; n=1 without CT data) of patients had liver metastases (24% [8/33] had >10 liver lesions), and 15% (5/33) had lung metastases (Figure 1B; Supplementary Table 1). Prior to first cfDNA collection, 47% of patients had received ≥1 line of AR-targeted therapy and 17% had received ≥1 line of taxane chemotherapy.

We leveraged a multi-omic profiling approach to assess epigenotypes from a single blood sample (Figure 1A). Plasma cfDNA was subjected to H3K4me2 cfChIP-seq using experimentally validated antibodies (21,28,29) to capture active gene promoters and distal cis-regulatory elements, plus (in select samples) several additional histone markers with functional relevance in human cell populations (Figure 1A; Supplementary Tables 2–3; Supplementary Figure 2A–B). Unlike prior cfChIP-seq studies (22,23) focusing on H3K4me3 (which captures active promoters) or H3K27ac (captures active promoters and enhancers but can be more labile compared to methylation marks, diminishing signal-to-noise and replicate reproducibility (42,43)), we pursued H3K4me2 to enable streamlined synchronous analysis of both promoter and enhancer activity using a single assay (32,44–52) (N.B. an identical rationale for solely analyzing H3K4me2 has been applied in previous human model system studies (53)) (Supplementary Figure 4). Quantitating enhancer activity is crucial as the enhancer landscape more strongly influences lineage identity than promoters and represents the primary regulatory element engaged by established prostate- and NEPC-specific transcription factors (e.g. AR, FOXA1, HOXB13, ASCL1, NEUROD1, NANOG) (10,16,54). The utility of H3K4me2 for promoter profiling is also supported by prior systemic functional mapping of 11 histone modifications, indicating that TSS-proximal H3K4me2/3 are approximately equivalently predictive of transcription compared to activating acetylation markers (H3K9ac and H3K27ac) (32). Samples also underwent deep targeted sequencing of cfDNA (median 1402× depth) and matched white blood cell (WBC) DNA to inform on cancer driver alterations and simultaneously resolve germline or clonal hematopoietic variants (Supplementary Tables 4–5). Our targeted panel captures 9087 genome-wide heterozygous germline single nucleotide polymorphisms (SNPs)—which together with a low-pass (median 0.96×) whole-genome input control for cfChIP-seq peak-calling—facilitated resolution of broad chromosomal instability and whole-genome duplication status (Figure 1B) as well as orthogonal inference of ctDNA%. Genotypic features from bulk ctDNA reflected clinically-aggressive mCRPC. Plasticity gatekeepers TP53, RB1, and PTEN were frequently disrupted by somatic alterations (Figure 1B), and the AR gene region (including its upstream enhancer) exhibited a broad range of activating alterations including frequent amplification and complex structural variants (Figure 1C) (18,38,55). 35% (15/43) of evaluable mPCa had samples harboring whole-genome duplication (WGD), and 4 had chromosomal-arm hallmarks suggesting chromothripsis.

cfChIP-seq profiles displayed expected histone modification spatial patterns: both H3K4me2 and H3K4me3 canonically marked active promoters, although H3K4me2 also showed clear signal across gene bodies and other non-TSS locations (Figure 1D–E). By contrast, H3K36me3 density accumulated toward the 3’-end of genes consistent with active transcriptional elongation. Repressive chromatin is marked by H3K27me3 and was observed at inactive promoters (Figure 1D,F). Incorporating all profiled markers, the AR TSS and upstream enhancer were strongly enriched for activating epigenomic signals (including H3K4me2) in mCRPC ctDNA—mirroring the LNCaP cell line positive control—yet harbored comparatively weak signal in non-prostate lineages (i.e. bladder cancer and ctDNA-negative samples) (Figure 1D) (56). In 17 mPCa samples with dual H3K4me2/3 profiling, H3K4me2 (but not H3K4me3) was strongly enriched across the AR enhancer, corroborating the utility of H3K4me2 for prostate-relevant enhancer quantification (Supplementary Figure 5). Expectedly, high ctDNA% cfChIP-seq had markedly higher concordance with primary prostate cancer tissue ChIP-seq from Pomerantz et al. 2020 than ctDNA negative cfChIP-seq (Supplementary Figure 6) ((2,18,33)).

We explored the relationship between TSS histone modification density in a ctDNA-negative sample and RNA expression in healthy blood (37). Activating H3K4me2/3 counts were expectedly positively correlated with whole blood expression while repressive marker H3K27me3 was negatively correlated, although the effect size for all markers was modest (absolute Spearman R<0.6) (Figure 1F). Interestingly, H3K4me2/3 fragment count density appeared bimodal (i.e. “low” or “high”) across the entire range of expression, with the proportion of methylated TSSs increasing as a function of expression (in contrast to a continuous, linear monotonic relationship which was the assumption in early cfChIP-seq studies) (Figure 1F) (23). Genes bivalently marked by H3K4me2 and H3K27me3 expectedly had lower expression than those with just H3K4me2 (Figure 1F).

cfDNA histone modifications reflect prostate lineage-specific biology

Bulk cfDNA represents admixed tumor-derived and normal cfDNA from non-tumor lineages (21). To determine whether plasma H3K4me2 can capture prostate cancer-specific biology, we investigated differences between high and low-ctDNA% samples, hypothesizing that established prostate-(cancer) specific and hematopoietic-specific factors would be strongly correlated and anticorrelated with ctDNA%, respectively.

We first examined the spatial distribution of H3K4me2 across genes highly expressed in prostate cancer (e.g. HOXB13, FOXA1, KLK3 [encodes PSA, prostate specific antigen], FOLH1 [encodes PSMA, prostate specific membrane antigen]) or blood (Figure 2A; Supplementary Figure 7). H3K4me2 density in prostate genes was markedly higher in ctDNA-positive samples than in ctDNA-negative, healthy controls, and WBC samples, whereas hematopoietic lineage-specific genes (e.g. CST7) showed an inverse relationship (Figure 2A). We observed that ctDNA%-correlated H3K4me2 signal could fall anywhere within gene bodies (i.e. TSS to transcription terminus site) or flanking regions and was not necessarily strongest at the TSS. For example, we observed differential H3K4me2 enrichment (of varying magnitude) in adenocarcinoma versus NEPC across multiple known intragenic and upstream FOLH1 cis-regulatory elements (Supplementary Figure 7) (57). Therefore, we next correlated ctDNA% to H3K4me2 density separately for TSSs (Figure 2B–C) or gene bodies (Figure 2D–E) across 17,774 genes (Supplementary Figure 2C). Since somatic copy alterations modulate the ratio of normal to tumor-derived cfDNA at individual loci, we utilized local gene-specific ctDNA fractions. Genes most positively correlated with local ctDNA% appeared strongly enriched for known prostate cancer genes and associated with pathways that are dysregulated in prostate cancer (Supplementary Table 11), whereas genes most negatively correlated with local ctDNA% were analogously enriched for hematopoietic signals (Figure 2B–E). Highly correlated genes aligned with expected RNA abundance from mCRPC and blood cells, modeled as a linear combination of the two tissue types weighted by gene-specific ctDNA% (Figure 2C,E). These observations extended to H3K4me2 counts at the binding sites of 458 transcription factors (TF) (Supplementary Tables 6–7), similarly indicating that key prostate lineage TFs (e.g. AR, FOXA1) were most strongly linked to ctDNA% (Figure 2F–G). Utilizing genome-wide H3K4me2 fragment counts, principal component analysis revealed a strong correlation between principal components 1 and 2 and sample ctDNA% (Figure 2H). Collectively, these data indicate that histone features specific to prostate cancer cells are evident in bulk cfDNA from patients with mPCa. Importantly, cancer signal strength is strongly influenced by ctDNA%, reinforcing the need to account for ctDNA% as a variable when considering apparent differences between samples.

Figure 2. Epigenomic footprints of cancer and normal cells contributing to bulk cfDNA.

(A) Regional H3K4me2 fragment counts across 4 lineage-specific genes in cfDNA and WBC DNA samples. Samples (rows) are sorted by local ctDNA%. B-G: Examples of the individual linear models of H3K4me2 fragment counts versus local ctDNA% (each dot is a cfDNA sample) [left column]. Adjacent plots [right column] aggregate these per-factor linear model statistics across 17,487 transcription start sites (TSS) (C) and 8071 gene bodies (each dot is a gene) (E) and published binding sites of 458 TFs (each dot is a TF) (G); x-axis values in C,E,G represent the log ratio of pure ctDNA and pure normal cfDNA fragment counts inferred from each per-factor linear model fit (i.e. the y-intercepts at boundary conditions of 100% and 0% local ctDNA fraction, respectively). (B) H3K4me2 fragment count in a 2kb neighborhood around CST7 TSS. (D) H3K4me2 fragment count across the HOXB13 gene body. (F) Mean H3K4me2 fragment count in a 2kb neighborhood of n=37,386 FOXA1 TF binding sites. In C and E, color is used to indicate the top and bottom 20% of genes ranked by their expression ratio between mCRPC and whole blood. Kernel density estimates of fragment count log ratios (i.e. x-axis values) are shown above the scatter plots (only includes genes with significant correlations [p<0.01] from linear model fitting). In G, different published TF ChIP-seq experiments are indicated by color. Kernel density estimates of TF fragment count log ratios are similarly shown. (H) Sample ctDNA% is strongly correlated with principal components one and two of H3K4me2 fragment counts across genome-wide consensus peaks. (I) Average spatial H3K4me2 enrichment distribution across 18 sets of open chromatin regions specific to distinct cancers, visualized in patient P12 (mPCa) and P8 (bladder cancer). In the mPCa sample, the most pronounced enrichment is observed in the prostate cancer trace, while in the bladder cancer sample, the predominant enrichment is observed in the bladder cancer trace. (J) Prostate and bladder cancer cfDNA has higher H3K4me2 fragment counts in prostate- and bladder-specific open chromatin compared to any other cancer type, respectively. (K) Spatial H3K4me2 density across 18 cancer-lineage specific accessible chromatin regions, demonstrating strong colon and prostate cancer signal in patient P30. (L) CT imaging showing locally recurrent prostate cancer with direct invasion of the adjacent rectal cavity in patient P30.

To further explore the dominant phenotypic/tissue lineages in our cfDNA samples, we quantified H3K4me2 enrichment across 18 different cancer lineage-specific accessible chromatin regions, predominantly capturing distal elements (e.g. enhancers) and promoter regions (36). ctDNA-positive mPCa samples exhibited higher H3K4me2 intensity within prostate cancer-specific open chromatin regions compared to any other cancer-specific accessibility landscape (p=5×10⁻⁹, Mann-Whitney U [MWU] test), whereas ctDNA-negative, healthy controls, and WBC samples had uniformly low H3K4me2 intensity (Figure 2I–J; Supplementary Figure 3). Underlying anatomic identity dominates the global regulatory architecture of malignant cells from the same lineage (58). Fascinatingly, one patient (P1) with innumerable lung lesions had exceptionally high H3K4me2 intensity in lung adenocarcinoma accessible regions (Supplementary Figure 3), while another patient (P30) with rectal invasion from a local recurrence within the prostate resection bed harbored strong colorectal cancer signal (Supplementary Figure 3, Figure 2K–L). Although not all documented sites of patient metastasis were strongly evident in the accessibility landscape (and we did not test these relationships systematically), these anecdotes raise the possibility that plasma cfChIP-seq can reveal patient-specific variation in prostate cancer phenotype while simultaneously detecting signals from metastasis-induced organ destruction (59).

Plasma cfChIP-seq reveals biological distinctions between clinical subgroups

Distinct patterns of metastases correlate with differential mCRPC prognosis (60), although the molecular determinants of organotropism and associated outcomes are incompletely understood. We tested whether plasma cfChIP-seq can unveil molecular distinctions between patient subgroups dichotomized by radiographic, laboratory, and genomic features. Recognizing that the observed histone modification signal originates from both ctDNA and non-tumor cfDNA—which contribute proportionally to the bulk total count of DNA fragments recovered from any given genomic region—we modeled cancer population H3K4me2 fragment counts of 17,774 genes and the binding sites of 458 TFs using multiple linear regression with local ctDNA% interaction terms, and tested significance of group differences with F-test (Methods, Supplementary Figure 2D).

Most mCRPC are thought to fall on a continuum of androgen reliance versus neuroendocrine feature dominance, possibly underlying variability in response to standard AR-targeted therapies (13,61). Current clinical tools are largely insufficient for categorizing patients along this clinically-relevant continuum (26). Consistent reciprocal enrichment/depletion of AR and neuroendocrine-factors was strongly evident from our plasma cfChIP-seq data (Figure 3A–I; Supplementary Figures 8–11): patients with biopsy-confirmed NEPC had low AR activity and were strongly differentially enriched for lineage reprogramming and pluripotency factors with established relevance in NEPC model systems (e.g. EZH2, NANOG, POU5F1) (Figure 3C; Supplementary Figure 9) (12,13,55,62–64). Fascinatingly, similar sets of NEPC-related factors were among the most differentially upregulated genes/TFs across separate analyses of patients with high liver lesion burden (Figure 3A–B), low PSA (Figure 3D), and RB1 deletions (consistent with prior model system data implicating RB1 as a lineage gatekeeper (55,65)) (Figure 3C; Supplementary Figure 8), implying a degree of epigenotype convergence (and lineage plasticity) even in patients without biopsy-confirmed NEPC. Conversely, patients with high burden bone metastases or high PSA were separately enriched for factors consistent with sustained androgen addiction (e.g. KLK3, AR [binding sites derived from primary prostate cancer tissue], FOXA1 and EP300 [both binding sites derived from estrogen-sensitive MCF-7 cells]) and were depleted for NEPC-related signals (Figure 3B,D) (2,14,66). Patients with high burden bone lesions (which were predominantly osteoblastic on CT) also harbored outlier enrichment of several Wnt-pathway TFs, which in prostate and breast cancer is implicated in osteo-mimicry (67). We observed an especially strong positive relationship between KLK3 TSS H3K4me2 intensity and time-matched serum PSA, offering proteomic validation that cfChIP-seq can measure variables of disease burden and androgen dependence (23) (Figure 3D). As a negative control, bladder cancer samples lacked H3K4me2 at KLK3 and AR TF binding sites (consistent with minimal AR signaling in bladder cancer). Finally, genes with elevated H3K4me2 in prostate compared to bladder cancer were associated with higher relative expression in prostate cancer (and vice versa). Interestingly, APOBEC3D (a known driver of mutagenesis in bladder urothelium (68)) and UPK2 (a highly specific urothelial diagnostic marker (69)) appeared enriched in bladder cancer (Figure 3E).

Figure 3. Epigenomic correlates of clinically-stratified metastatic prostate cancer.

(A) Mean H3K4me2 fragment counts of AR (n=3,223) and NANOG (n=11,611) TF binding sites versus average local ctDNA% (each sample is a dot; one sample plotted per patient). Only samples evaluable for both liver and bone metastases are shown in the scatter plots (with two exceptions, annotated). (B) Volcano plots comparing fragment count between cfDNA samples in patients with synchronous high- versus low-burden liver (left) or bone metastases (right), focusing on the binding sites of 458 TFs (each dot is a TF, notable outliers in B and D are highlighted). Volcano plot x-axis values in B,D,E represent the log ratio of pure ctDNA to pure normal cfDNA fragment counts between the two sample groups, as inferred from each per-factor linear model (methodology and figure walkthrough shown in Supplementary Figure 1D). Y-axis shows interaction F-test p-values for the sample group term in the linear model. (C) Log ratios and F-test p-values for the top 8 differentially enriched TFs across 6 categorical comparisons—the top and bottom 5 TFs are displayed (among the set of TFs with F-test p<0.05 [no multiple hypothesis correction]). Number of samples in each clinico-genomic category and the overlap between categories is shown left (only samples with >10% ctDNA are included). (D) KLK3 TSS H3K4me2 fragment count, local ctDNA% and time-matched serum PSA are strongly correlated. Box plots are dichotomized by median PSA and include only samples with >50% ctDNA at KLK3 locus. Fitted linear model includes a continuous term for log(PSA+1). Two model instantiations are illustrated representing PSA=0 and PSA=1000. (E) Gene body H3K4me2 fragment counts between bladder and prostate cancer. Volcano plot data points are color-coded for top and bottom 20% gene expression ratio between prostate and bladder cancer. Kernel density estimates of fragment count log ratios only include genes with p<0.01. (F) NEPC- and AR-related motifs enrichment in mPCa patients stratified by clinical subgroup and sample ctDNA%. MWU tests were performed on motif enrichment values. (G) AR- and NEPC-specific TF motif enrichments demonstrate contrasting correlations with PSA. (H) Negative correlation between the AR- and NEPC-TF motif enrichment in ctDNA-positive mPCa. (I) Average H3K4me2 signal in open prostate-cancer chromatin regions is higher in patients with prostate adenocarcinoma than NEPC (MWU p=0.051, calculated on fragment counts of ctDNA positive samples). A-I Only the highest ctDNA% sample per patient is represented.

To orthogonally validate these links between mPCa clinical features and AR-dependency, we next interrogated H3K4me2 counts at predicted TF binding motifs throughout the genome, focusing on the core TF regulators of recently proposed mCRPC epigenomic subtypes (10). Consistent with our aforementioned interaction tests, TF binding motifs associated with the CRPC-NE and CRPC-AR epigenomic subtypes were differentially enriched and depleted (respectively) in patients with high burden liver metastases, relative to those with low burden or no liver lesions (Figure 3F). Increasing PSA was positively correlated with CRPC-AR TF motif enrichment (Spearman correlation: 0.41, p<0.004) and negatively correlated with CRPC-NE TF motif enrichment (Spearman correlation: −0.36, p<0.015) (Figure 3G). CRPC-AR and CRPC-NE motif enrichment expectedly appeared mutually exclusive (Figure 3H). Intriguingly, samples from patients with neuroendocrine features had markedly lower H3K4me2 intensity in prostate cancer accessible chromatin regions compared to the cohort median (p<0.051, MWU) (Figure 3I), consistent with prior observations of prostate adenocarcinoma and NEPC harboring distinct chromatin architectures in model systems (16). Highlighting the potential for plasma cfChIP-seq to identify temporal changes in disease phenotype, we observed clear epigenomic switching towards NEPC in one patient that provided two plasma samples 16 months apart (Figure 4A). Here, H3K4me2 counts at TF motifs supported the (tissue biopsy confirmed) transition to ASCL1-driven small cell phenotype from previously AR-dominant disease (15,16). Interestingly, this occurred post-exposure to docetaxel and lutetium-PSMA radioligand treatment, agents not traditionally implicated in inducing treatment-emergent NEPC.

Figure 4. Temporal lineage variability inferred via serial plasma cfChIP-seq.

(A) Left: Treatment history, PSA timeline, and radiographic imaging for Patient P20, a 76-year-old with treatment-refractory mCRPC with bone and lymph node disease at time of first plasma cfDNA collection, without evidence of visceral metastases. Approximately 12–16 months after initial plasma cfDNA collection, following deep biochemical responses to both docetaxel and Lutetium-PSMA radioligand therapy, he experienced dramatic disease progression, manifested by near-complete infiltration of liver parenchyma with metastatic disease. A second plasma cfDNA sample was obtained, and peri-collection liver biopsy confirmed emergence of therapy-induced small cell carcinoma (with concurrent adenocarcinoma), supported by positive immunohistochemical staining for classic neuroendocrine markers chromogranin and synaptophysin and low serum PSA (violin plot, right). Despite treatment with platinum doublet chemotherapy, he died shortly after due to multi-organ system failure including malignant obstructive uropathy—consistent with outlier elevated prognostic markers LDH and ALP (in-set violin plot; dots indicates patient P20 clinical marker values [post-lutetium-PSMA] relative to the whole-cohort distribution measured at first ctDNA collection) and strong H3K4me2 open chromatin enrichment across multiple distinct non-prostatic tissue lineages (Supplementary Figure 6C). Comparison of serial plasma cfChIP-seq showed enrichment of multiple transcription factor motifs associated with NEPC transdifferentiation, most notably in ASCL1, but also a more modest signal increase in NeuroD1 (15,16), coinciding with the development of fulminant liver metastases. Contrastingly, after accounting for differences in sample ctDNA%, no overt differences in copy number architecture were apparent between cfDNA collection timepoints. (B) Left: Treatment history, radiographic imaging and histopathology for patient P14. Pre-docetaxel CT imaging revealed widespread metastatic disease, including bulky lung (red arrow) and pelvic (pink) lesions. Clinical deterioration coincided with the development of multiple new brain metastases (blue). Metastatic scapula biopsy H&E revealed a poorly-differentiated malignancy with extensive tumor necrosis, with no resemblance to the primary prostate adenocarcinoma biopsy taken 55 months prior (Supplementary Figure 11). Top right: H3K4me2 cfChIP-seq and genome-wide copy number plots comparing pre-docetaxel and on-treatment cfDNA samples. Radar plots show TF motif enrichment −log(p). Line plots show spatial H3K4me2 distribution in open chromatin regions specific to different cancer types. Bottom right: Whole genome copy number profiles of the two cfDNA samples of P14. VAF scatterplots comparing the second cfDNA sample to melanoma in situ biopsy and the first cfDNA sample. All mutations called in any of the three samples are shown. Abbreviations: ADT, androgen deprivation therapy; NTD, (AR) N-terminal domain; PSA, prostate specific antigen; CT, computed tomography; H&E, hematoxylin and eosin; lpWGS, low-pass whole-genome sequencing.

Advanced prostate cancer primarily affects an elderly demographic at heightened risk for other primary malignancies with metastatic patterns similar to those in prostate cancer. In late-stage disease, for which there are an expanding number of active therapies with proven survival advantage, identifying occult concurrent non-prostate advanced malignancies via ctDNA may alter optimal patient care. This was exemplified by case P14 with aggressive mCRPC and disease (epi-)genomic features strongly consistent with prostate cancer in his initial blood collection (Figure 4B). Despite a deep and sustained biochemical response to docetaxel, the patient deteriorated clinically secondary to hemiparesis and innumerable new bihemispheric brain metastases. Although this clinical presentation led to the assumption of de-differentiated, non-PSA-secreting prostate cancer (e.g. NEPC or double-negative mCRPC), the second blood collection showed high tumor mutation burden (19.3 mutations/Mb) and shared no genomic alterations with pre-docetaxel ctDNA. H3K4me2 intensity across 18 cancer-specific accessible chromatin regions revealed a strong relative shift towards melanoma and diminished prostate identity in the progression ctDNA. A metastatic shoulder biopsy taken before docetaxel initiation tested negative (by immunohistochemistry) for prostate and neuroendocrine lineage markers (Supplementary Figure 12). Intriguingly, the patient had an excised melanoma in situ from five years earlier at time of metastatic prostate cancer diagnosis. Mutational overlap between the melanoma tissue and second cfDNA sample confirmed shared ancestry.

Discussion

In this proof-of-concept study, we develop a new plasma cfChIP-seq analysis framework and systematically interrogate links between plasma cfDNA histone markers and synchronous radiographic and clinical features in lethal prostate cancer. Our data implicate epigenomic states along the neuroendocrine:androgen-reliance axis as underlying common differences in mCRPC clinical presentations, including in patients without histologically-confirmed NEPC. Additionally, we highlight the potential for cfChIP-seq to differentiate superposed lineages in bulk cfDNA—such as DNA originating from other primary malignancies and metastasis-induced tissue pathophysiology. Collectively, these data authenticate plasma cfChIP-seq as a minimally-invasive in vivo biological discovery tool and nominate potential opportunities for this technique to influence individual patient management.

All epigenomic features recovered from bulk cfDNA are profoundly modulated by ctDNA% (6,8,9,19). While variable tumor purity also confounds ChIP-seq analysis of tumor tissue, in cfDNA tumor purity is generally substantially lower and interpatient variance is higher even within clinically homogenous cohorts controlling for tumor type, stage, and timing of blood collection. However, unlike tissue tumor purity, ctDNA% is strongly influenced by multiple clinically relevant features including metastatic burden and pattern of organ involvement, tumor phenotype, and patient prognosis (40). Therefore, failing to account for ctDNA% means that apparent epigenomic differences between individual genes or patient subsets (i.e. both intra- and inter-sample comparisons) may be entirely driven by differences in ctDNA%, owing to the distinct epigenomic properties of tumor and non-tumor cfDNA. The confounding nature of ctDNA% was evident in our cohort despite strategic enrichment for high ctDNA% samples (to boost mCRPC-specificity), where the epigenomic footprints of most genes and TFs correlated with ctDNA% (Figure 2A–G). Existing ChIP-seq analysis paradigms were largely developed for pure cancer samples derived from model systems, meaning that the challenge of heterogeneous ctDNA% for clinical profiling is relatively new. Nevertheless, we demonstrate that plasma cfChIP-seq is a powerful tool for generating tumor-specific biological and clinical insights, contingent on careful adjustment for ctDNA%. While supervised adjustment for ctDNA% can potentially enhance any cfChIP-seq analysis, it is especially important for optimizing sensitivity and specificity of ctDNA-based classifiers predicated on few features (e.g. select TSSs, genes, or distal elements) where ctDNA% cannot be easily learned from input features alone (Supplementary Figure 2E). Our work provides a new statistical blueprint to control for variable ctDNA% when applying cfChIP-seq to large cohorts, by modeling patient-group comparisons as interaction tests with ctDNA%, followed by empirical comparison to a background distribution (i.e. all genes or TFs). Existing cfChIP-seq workflows can readily incorporate our ctDNA%-aware approach through estimating copy numbers and genome equivalent ctDNA% (e.g. by applying standard tools like ichorCNA (31)) from the input low-pass WGS data typically already generated for conventional ChIP-seq signal normalization. Adjustment for ctDNA% is already standard practice for interpreting ctDNA genomic analyses (e.g. to limit false negatives)—yet despite the equivalent (or higher) relevance of ctDNA% for understanding cfDNA epigenomic signals—is not incorporated into existing cfDNA epigenomic workflows. Importantly, our methodological framework can also be applied to other emerging ctDNA epigenomic profiling modalities (e.g. 5(hydroxy)methylcytosine sequencing). To maximize tumor phenotype resolution, future biological discovery research utilizing plasma cfChIP-seq should prioritize patients with high ctDNA%—accepting the limitation of enriching for poorer prognosis—while simultaneously utilizing ctDNA negative samples to control for the contribution of non-tumor cfDNA. High ctDNA (above ~20%) is relatively common in many common metastatic cancers (e.g. breast, bladder, lung (39,70)) and is typically conserved across serial progressions (40), indicating significant discovery potential for this technique beyond mCRPC (wherein ~25–30% of unselected first-line patients have ctDNA>20%).

Our study demonstrates that plasma cfChIP-seq can reveal epigenomic distinctions between clinically-stratified patients. A key finding is the striking convergence on the AR:neuroendocrine axis in underlying common clinical features, as revealed through a systematic search for differentially enriched genes/TFs across patient subgroups (Figure 3). This suggests that cell-free histone markers have potential utility for measuring tumor androgen reliance (e.g. to predict androgen receptor pathway inhibitor response). Importantly, neuroendocrine epigenomic hallmarks were evident even in patients without histologically-confirmed NEPC, implying that dedifferentiated phenotypes may be underappreciated in advanced mCRPC. The observation that radiographically stratified patients were enriched for different features more generally highlights the potential for plasma cfChIP-seq to nominate molecular determinants of organotropism—contingent on sufficiently detailed radiographic annotation. PSA was strongly correlated with KLK3 H3K4me2 intensity (Figure 3D), foreshadowing opportunities for cfChIP-seq to provide a functional readout of other clinically-relevant disease markers (23). For example, plasma cfChIP-seq may enable in vivo validation of emerging cell-surface targets to facilitate development of new therapeutic and functional imaging approaches. Future studies should investigate whether cfDNA histone markers in FOLH1 (encodes PSMA) (57,71) (Supplementary Figure 7) or DLL3 predict radiotracer avidity in prostate cancer—as a strategy to either augment or circumvent existing targeted radioimaging-based treatment selection (e.g. for lutetium-PSMA). This may be particularly advantageous for predicting vulnerability to therapies (e.g. sacituzumab govitecan in breast cancer) targeting cell surface markers (TROP-2) that are currently only evaluable via metastatic tissue analysis.

Multiple non-prostate populations were identifiable in bulk cfDNA—including from normal tissues and other primary malignancies—indicating opportunities to inform on additional pathology relevant for patient oncological care. Our data reaffirm prior studies demonstrating a hematological origin of most non-tumor cfDNA (21). Fascinatingly, non-tumor cfDNA in some patients also harbored clear epigenomic footprints of locally-invasive malignant infiltration (e.g. the lung and rectal wall) (Figure 2K,L; Supplementary Figure 3), indicating hypothetical utility for cfChIP-seq to forecast cancer-related organ damage ahead of clinical symptoms, or monitor for real-time functional improvement after therapeutic intervention (e.g. via metastasis directed therapy or systemic corticosteroid administration) (59). These theoretical applications of plasma cfChIP-seq will require highly sensitive and specific assays, as well as deeper understanding of microenvironmental or other mechanistic factors influencing local cfDNA release—evident from our study in that not all metastatic lesions had accompanying normal tissue H3K4me2 footprints in cfDNA. In the context of monitoring for treatment-related adverse events, it is plausible that analytic sensitivity may be highest during active effective therapy when ctDNA is suppressed.

Our study has several limitations. First—given the priority of hypothesis generation and initial exploration of plasma cfChIP-seq’s biological discovery potential (requiring high ctDNA%)—we did not formally benchmark analytical performance across the full spectrum of ctDNA% observed in unselected populations. Although we observed distinct H3K4me2 patterns between ctDNA-positive (ctDNA > 20%) and -negative (ctDNA < 1%) samples, the complexity of H3K4me2 biology (e.g. varying factor-specific magnitude of differential enrichment in ctDNA versus normal cfDNA) and lack of consensus on quantification approach (e.g. analyzing TSS, gene body, and/or other regulatory elements) impedes obvious metrological assessment. Enriching for patients with high ctDNA% may also bias our cohort toward more extreme or nonrepresentative phenotypes relative to the general mCRPC population. Second, we were unable to differentiate regulatory states defined by combinatorial histone modification (e.g., primed [H3K4me2 without H3K27ac] versus poised [H3K4me2+H3K27me3] versus active [H3K4me2+H3K27ac] enhancers) (72). Relatedly, directly comparing H3K4me2 fragment counts between genes as a surrogate for differential activity is confounded by the multitude of other regulatory mechanisms which were not investigated (e.g. post-transcriptional and peri-translational regulation), ideally requiring matched RNA or protein for functional validation. The potential for H3K4me2 antibody methylform nonspecificity may also result in signal leakage between functionally disparate regulatory regions (e.g. promoters versus enhancers) (28). Third, leveraging cancer-associated open chromatin landscapes to infer normal tissue identity in bulk cfDNA (via H3K4me2 enrichment) may lack comparable sensitivity/specificity across tissue types, especially in scenarios where tumors harbor a distinct cell-of-origin or undergo substantial chromatin reorganization during malignant transformation. Fourth, public ENCODE TF-binding site ChIP-seq experiments used were mostly in non-prostate (cancer) specimens (32), potentially restricting generalization to our cfDNA samples due to differences in lineage-specific cistromes. We tolerated this and other shortcomings of ENCODE data (namely the absence of key NEPC-defining TFs ASCL1 and NEUROD1 (15,16)) in exchange for the advantage of uniform processing and rigorous quality control. Finally, we did not systematically test the ability for plasma cfChIP-seq to detect concurrent non-prostatic malignancies due to their relative rarity and likely frequency of underdiagnosis. Nevertheless, robust detection of diverse mCRPC phenotypes, admixed normal organ footprints, and epigenomic distinctions between prostate and bladder cancer (i.e. the most common secondary malignancy in men receiving primary prostate radiotherapy (73)) support the potential for such application.

Ultimately our data demonstrate that plasma cfChIP-seq adds to existing nascent cfDNA profiling modalities capable of quantifying lineage heterogeneity in bulk cfDNA, alongside fragment-based footprinting (18–20), and targeted and whole-genome 5(hydroxy)methylcytosine sequencing (4–6,8,9). However the sensitivity and specificity of these techniques for detecting lineage variability has not been directly compared. To complicate matters further, mCRPC phenotype is increasingly appreciated to be highly multidimensional (e.g. dedifferentiation as a continuous variable, existence of double-negative and amphicrine states, potential WNT and stem-cell phenotypes, subtype co-occurrence, and variable subtype burden) (7,10,13,61). Understanding differences in the precise biological implications of DNA (hydroxy)methylation, nucleosome footprints, and histone modifications will therefore be a key research priority. Plasma cfChIP-seq theoretically allows simultaneous interrogation of multiple histone markers with distinct regulatory functions, potentially enabling more multiplexed insight into tumor phenotype and clinical relevance (versus other techniques). Finally, the composition of normal cfDNA was previously largely irrelevant for ctDNA genotyping (except for clonal hematopoiesis), but for epigenomic profiling, methodological awareness of heterogeneous lineage composition in normal cfDNA will be key for true tumor resolution.

Statement of significance.

Plasma cfChIP-seq enables phenotypic dissection of lethal prostate cancer and is a practical tool for biomarker discovery, while overcoming prior limitations of access to relevant tissue and reliance on model systems.