Explainable machine learning identifies a one-carbon–epigenetic axis associated with sperm DNA fragmentation in reproductive-age men with cancer: a retrospective translational cohort study

Background Cancer can perturb systemic metabolism, redox homeostasis, and germ-cell genomic integrity before gonadotoxic therapy. Homocysteine (Hcy), a central intermediate of one-carbon metabolism, links methyl-donor balance, oxidative stress, and DNA methylation. We investigated whether serum Hcy was associated with sperm DNA fragmentation in reproductive-age men with cancer and whether explainable machine learning could integrate clinical, semen, and molecular features to characterize a candidate one-carbon–redox–epigenetic axis associated with high DNA fragmentation index (DFI). Methods This retrospective translational cohort included men referred for fertility preservation before systemic therapy or radiotherapy. The primary outcome was high sperm DFI (≥30%). Multivariable logistic regression and three machine-learning algorithms were evaluated using a 70% development/30% hold-out split. An archived-sample substudy assessed oxidative stress, global sperm DNA 5-methylcytosine, and MTHFR- and DNMT3A-related expression as candidate readouts of one-carbon and epigenetic regulation. Results Among 286 men, high DFI occurred in 26.2%. Serum Hcy correlated with DFI (Spearman ρ = 0.476, P < 0.001). After adjustment for age, BMI, smoking, varicocele, recent fever, and cancer type, each 5-µmol/L increase in Hcy was associated with higher odds of high DFI (adjusted OR 7.10, 95% CI 3.69–13.67). Logistic regression showed the highest hold-out discrimination (AUC 0.792), and permutation importance identified Hcy as the leading predictor. In 72 archived samples, higher Hcy and DFI were associated with greater oxidative stress, lower global 5-methylcytosine, and lower DNMT3A-related expression; MTHFR-related expression was inversely associated with Hcy. Conclusion In men with cancer before gonadotoxic therapy, higher serum Hcy was independently associated with sperm DNA fragmentation and contributed substantially to explainable prediction models. The molecular substudy supports, but does not establish, a candidate one-carbon–redox–epigenetic regulatory axis linking systemic metabolic state to germ-cell chromatin vulnerability. These findings provide a clinically anchored framework for future validation using high-resolution methylome, chromatin-accessibility, transcriptomic, and regulatory-network analyses.