Non-invasive prostate cancer subtyping using 5hmC-sequencing of cell-free DNA.

Abstract

242 Background: Optimizing treatment for metastatic castration-resistant prostate cancer (mCRPC) requires knowledge of different lineage subtypes such as androgen receptor (AR)- positive prostate cancer (ARPC), neuroendocrine prostate cancer (NEPC) and double-negative prostate cancer (DNPC). However, metastasis biopsies are not always available. Epigenetic and genetic characterization of cell-free DNA (cfDNA) has the potential to non-invasively classify tumor subtypes to help guide precision oncology and monitor tumor evolution. In this study, we sought to assess whether using cfDNA 5-hydroxymethylcytosine sequencing (5hmC-seq) non-invasively predicts tumor subtypes and signaling defined in matched tissue from men with mCRPC. Methods: We performed 5hmC-seq and low-pass whole-genome sequencing (lpWGS) of plasma-derived cfDNA from 86 mCRPC patients with matched tumor tissue profiled with 5hmC-seq (N = 49) and RNA-seq (N = 86). We developed a 5hmC-seq based Random Forest-based hierarchical classifier to distinguish ARPC, NEPC and DNPC. The previously published Griffin and ctdPheno tools were used to classify samples based on lpWGS. The locked classifiers were evaluated in a separate cfDNA cohort from men with mCRPC (N = 64). Results: Samples with less than 10% circulating tumor DNA fraction (ct-fraction) were categorized as a good prognosis group, while the rest samples were used for classifier training and testing (ARPC (N = 22), NEPC (N = 6) and DNPC (N = 11)). The 5hmC-based Random Forest subtype classifier reached an AUC of 0.89 [95% CI 0.73-1.0] and 0.88 [95% CI 0.76-1.0] for NEPC and DNPC prediction in cross-validation, respectively. The 5hmC-based predicted subtype in the separate validation cohort was strongly prognostic for overall survival (OS); NEPC compared to ARPC HR = 17.3 [95% CI: 2.0 – 147.9], p = 9.1 x 10-3, and DNPC compared to ARPC HR = 34.7 [95% CI: 5.3 – 225.6], p = 2.1 x 10-4, adjusting for ct-fraction. Furthermore, high 5hmC-seq NE score in cfDNA predicted shorter OS (HR = 8.3 [95% CI: 1.1 – 61.8], p = 0.038), while high 5hmC-seq AR score predicted longer OS (HR = 0.26 [95% CI: 0.09 – 0.75], p = 0.013), adjusting for ct-fraction. Concurrently sequenced lpWGS demonstrated that the ratio of short (< 150 bp) vs. long (150 ~ 220 bp) fragments was higher in NEPC compared to other subtypes at the similar ct-fraction (ANOVA test P = 0.0017). ARPC showed low central coverage at AR binding sites, while NEPC exhibited low central coverage at ASCL1 binding sites. DNPC displayed no signal at either binding site. The predicted subtypes based on nucleosome positioning patterns from lpWGS were concordant with that predicted by 5hmC-seq in the separate cohort. Conclusions: cfDNA 5hmC-seq based classifiers successfully identified tumor phenotype, classified tumor subtypes concordant with matched tissue and were prognostic. This non-invasive approach enables sequential monitoring of lineage plasticity during mCRPC progression.