Non-invasive classification & molecular subtyping of mature lymphoid neoplasms by cell-free DNA profiling

Abstract

Abstract Background: Mature lymphoid neoplasms (MLN) are currently diagnosed by integrating histopathological, immunophenotypic, cytogenetic, and molecular profiles of tumor tissue biopsies (Campo, Blood 2022; Alaggio, Leuk 2022). While genetic subtyping of DLBCL, based on mutational profiles and copy number alterations, has further advanced this classification, many cases remain unclassified (Wright, Cancer Cell 2020; Chapuy, Blood 2025). This highlights the need for additional diagnostic approaches, such as gene expression profiling (GEP) and cytogenetic methods, to capture significant biological heterogeneity. GEP has proven particularly valuable in identifying not only cell-of-origin (COO), but also high-grade B-cell lymphomas (HGBCL), offering enhanced sensitivity and the ability to detect occult rearrangements using a “dark-zone signature” (DZsig, Ennishi, JCO 2018; Alduaij, Blood 2023). However, the required tissue biopsies for GEP are invasive, carry procedural risks, and may delay diagnosis. While multi-cancer early detection (MCED) liquid biopsy tests based on methylation can allow detection of diverse solid tumors, most cancers detected by MCED when screening otherwise healthy adults are MLNs (Schrag, Lancet 2023) that cannot currently be further resolved using these tests. To overcome these limitations, we explored plasma cell-free DNA (cfDNA) as a non-invasive alternative for classifying diverse MLNs, determining COO and genetic subtypes in DLBCL, identifying genetic subtypes in HL, and detecting HGBCL. Methods: We applied EPIC-Seq (Esfahani, Nat Biotech 2022) to infer tumor gene expression from cfDNA fragmentomic signals in blood plasma before therapy. We designed a targeted capture panel informed by prior lymphoid GEP studies (n=56 studies, >10,000 tumors), including B- and T-cell differentiation markers, canonical immunohistochemistry markers, and recurrent genomic mutations and fusions for tumor burden assessment via variant allele frequency (VAF). The final panel encompassed 1,986 transcription start sites from 1,676 genes and covered 2.6 MB of genomic space. We then profiled plasma cfDNA from 748 subjects, including 666 patients across 10 lymphoid neoplasms (BL, CLL, DLBCL, FL, HL, MCL, MM, PBMCL, PTCL, WM), along with 58 healthy controls and 24 subjects with Other non-lymphoid solid tumors (carcinomas of the lung, liver, and pancreas). Results: Recognizing the influence of tumor burden on inferred tumor gene expression from plasma cfDNA, we developed a machine learning model to predict tumor mutant variant allele fraction (VAF) from EPIC-Seq data and validated it against mutation-based tumor burden measurements (Pearson R=0.96). This established a low VAF background threshold distinguishing from healthy controls (1%), below which samples were excluded for classification purposes. We then trained a multi-histology classifier on 396 patients across 10 groups (BL, CLL, DLBCL, FL, HL, MCL, MM, PTCL, Healthy, Other). Independent validation on 169 held-out samples demonstrated high individual classification accuracies and an overall top 2 accuracy of 94%. Beyond histological classification, EPIC-Seq showed high discriminatory potential for molecular HL subtypes (H1/H2, Alig, Nature 2024; AUC>0.8, p<0.0001) and accurately identified DLBCL GCB vs ABC COO subtypes (AUC>0.8, p<0.001). For genetic subtyping of DLBCL (e.g. Wright, Cancer Cell 2020; Chapuy, Blood 2025), we trained a classifier on publicly available tumor RNA-Seq data and successfully applied it to our plasma EPIC-Seq data from DLBCL patients, observing a strong concordance between EPIC-Seq inferred molecular subtypes and LymphGen mutational classifications. Finally, we evaluated cell-free DZsig in 49 LBCL patients (28% harboring MYC+BCL2 double-hit by FISH), demonstrating significant discrimination between HGBCL-DH-BCL2 and DLBCL NOS (AUC=0.75, p=0.005). As expected, DZsig scores in BL, known to originate from the dark zone, did not significantly differ from HGBCL-DH-BCL2.Conclusion: Our study demonstrates clinical feasibility and high accuracy of EPIC-Seq-based cfDNA analysis for non-invasive classification and molecular subtyping of diverse mature lymphoid neoplasms. This approach effectively addresses clinically relevant diagnostic challenges, including COO determination, genetic subtyping in DLBCL and HL, identification of HGBCL-DH-BCL2, and other high risk LBCLs harboring dark-zone signature.