Machine learning of genome-wide 5-hydroxymethylcytosines in blood identifies prognostic models for survival outcomes in multiple myeloma

ABSTRACT: Background: Aberrant epigenetic modifications play an important role in the progression and aggressiveness of multiple myeloma (MM). The MM-specific hydroxymethylome has been associated with MM cell proliferation and a lower global 5-hydroxymethylcytosines (5hmC) level is associated with inferior overall survival of MM. To date, no study has quantitatively investigated genome-wide 5hmC signatures in circulating cell-free DNA (cfDNA) for its prognostic significance in MM. Methods: Genome-wide 5hmC profiles were obtained in 321 plasma cfDNA samples collected at the time of diagnosis from newly diagnosed patients with MM between 2010-17 at the University of Chicago Medical Center with a median follow-up of 70.48 months. Deaths were ascertained using the National Death Index. Treatment data (types and transplantation) and outcomes (response, progressive disease, relapse, etc.) were collected from electronic medical records. The Cox proportional hazards model was used to identify 5hmC features associated with overall survival (OS) and progression-free survival (PFS) associated with 5hmC features, followed by pathway analysis and modeling using the elastic net regularization. Results: Cellular deconvolution showed monocytes as the major source of DNA in cfDNA. The 5hmC profiles in cfDNA reflected the 5hmC in genomic DNA from CD138+ tumor cells from bone marrow and differed by validated clinical prognostic indices. Altered 5hmC features (gene bodies) involving calcium signaling and neuroactive ligand-receptor interaction were found to be associated with OS. A weighted prognostic-score (wp-score) model comprising 18 5hmC modified genes was trained using machine learning, showing a hazard ratio (HR) of 2.9 (95% confidence interval [CI]: 1.7-4.9; log-rank test p<0.0001) for OS and HR of 1.8 (CI: 1.3-2.5; p=0.00014) for PFS in the validation set. Time-dependent receiver operating characteristic (ROC) analysis showed that the wp-score maintained its predictive capacity for OS and PFS at 24, 48, and 72 months of follow-up. The 5hmC-based prognostic model also outperformed known prognostic factors for MM, including ISS stage, LDH level, and treatment types. Conclusions: Genome-wide 5hmC profiles in patient-derived cfDNA samples reflected specific epigenetic signatures associated with survival outcomes. We demonstrated the feasibility of using non-invasive 5hmC biomarkers for the prognosis of patients with MM.

ORGANISM(S): Homo sapiens

PROVIDER: GSE186351 | GEO | 2026/08/05

REPOSITORIES: GEO