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 DNA (cfDNA) in humans may enable the capture of disparate prostate cancer phenotypes. In this study, 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 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 clinicogenomic features: bone- versus liver-predominant disease, serum PSA, biopsy-confirmed histopathologic 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 use 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 insights into the adaptive molecular landscape of aggressive prostate cancer and endorse plasma cfDNA chromatin profiling as a biomarker source and biological discovery tool. Significance: Plasma cell-free chromatin immunoprecipitation sequencing 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.

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.

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.

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.

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.