Non-invasive detection and classification of lymphoma via cell-free DNA methylation profiling

Abstract Diagnosing lymphoma relies on invasive tissue biopsies, which can yield insufficient material for histopathological evaluation and carry a risk of complications. Cell-free DNA (cfDNA) analysis from plasma represents a promising alternative for non-invasive lymphoma diagnosis, as DNA methylation patterns are both highly cell type–specific and characteristically altered in malignancy. We analyzed cfDNA methylation in 265 plasma samples (165 pre-treatment samples from lymphoma patients: 71 DLBCL, 46 FL, 48 HL; 48 non-lymphoma/non-malignant controls; and 52 post-cycle 1 or end-of-treatment EOT samples from 15 DLBCL and 12 FL patients) using cell-free methylated DNA immunoprecipitation and sequencing (cfMeDIP-seq). Most pre-treatment samples (84.2%) were obtained at diagnosis, and a small number of samples before second-line treatment (15.8%). The pre-treatment cohort was split into discovery (n=142) and validation (n=71) sets for model development and testing. Differential methylation analysis identified 13,934 lymphoma-associated hypermethylated regions, which were used to train regularized binomial generalized linear models. Enrichment analyses revealed these regions overlapped significantly with CpG islands and H3K27me3-marked genes. In the validation cohort, the binary classification model distinguishing lymphoma from controls achieved an accuracy of 0.88, with a positive predictive value (PPV) and negative predictive value (NPV) of 0.88. Subtype-specific models were subsequently developed: the DLBCL vs. control model reached an AUC of 0.96 and accuracy of 0.87 (PPV = 0.91, NPV = 0.84); the FL vs. control model yielded an AUC of 0.82 and accuracy of 0.74 (PPV = 0.82, NPV = 0.69); and the HL vs. control model achieved an AUC of 0.99 and accuracy of 0.96 (PPV = 0.94, NPV = 0.98). Stage-stratified analysis showed high classification performance for both limited and advanced-stage disease. The lymphoma vs. control model achieved AUCs of 0.96 (advanced-stage) and 0.94 (limited-stage). For DLBCL, AUCs were 0.98 and 0.94; for FL, 0.88 and 0.70; and for HL, 0.99 and 0.97, respectively. A three-class model distinguishing controls, HL, and a combined DLBCL/FL group showed robust overall performance. HL classification achieved an AUC of 0.99 and accuracy of 0.89 (PPV = 0.91, NPV = 0.89); the DLBCL/FL group reached an AUC of 0.95 and accuracy of 0.86 (PPV = 0.92, NPV = 0.82); and control classification had an AUC of 0.95 and accuracy of 0.80 (PPV = 0.69, NPV = 0.87). A four-class model distinguishing HL, DLBCL, FL, and controls showed that HL remained the most accurately identified subtype (AUC = 0.99, accuracy = 0.89, PPV = 0.92, NPV = 0.89), followed by DLBCL (AUC = 0.89, accuracy = 0.83) and FL (AUC = 0.80, accuracy = 0.79). DLBCL samples misclassified as FL were enriched for GCB-type mutations in EZH2 and BCL2 and lacked ABC-associated mutations such as TBL1XR1, BTG1, CCND3, and PRDM1. We calculated cfDNA methylation scores by averaging normalized methylation levels across lymphoma-associated hypermethylated regions. These scores were significantly associated with LDH levels (DLBCL: R = 0.53, p = 2.1×10⁻⁶; FL: R = 0.51, p = 2.9×10⁻⁴), IPI in DLBCL (p = 0.0077), FLIPI in FL (p = 9.1×10⁻⁸), cfDNA tumor burden, and metabolic tumor volume from PET-CT. Methylation scores from plasma samples taken after the first immunochemotherapy cycle (15 DLBCL, 10 FL) and at EOT (15 DLBCL, 12 FL) tracked treatment response as conveyed by PET-CT or CT scans. Increasing scores were observed alongside radiographic progression in 2 patients, and a patient with complete radiological response but with a slow declining methylation score post-cycle 1 had early progression 2 months after EOT. Five additional patients with low EOT methylation scores and complete metabolic response experienced either relapse or transformation. Four progression-free patients showed partial radiological response, but had low methylation scores at EOT. To the best of our knowledge, this is the first study to apply cfMeDIP-seq to plasma samples from lymphoma patients. cfDNA methylation profiling offers a sensitive, minimally invasive approach for lymphoma detection and subtype classification, with high classification performance even in early-stage disease for DLBCL and HL. cfDNA methylation correlates with tumor burden and clinical risk, supporting its potential role as a biomarker for predicting treatment response.