Diffuse large B-cell lymphoma (DLBCL) is the most common malignancy of the lymphohematopoietic system in adults, accounting for approximately 35% of all non-Hodgkin lymphomas. Despite significant advances in immunochemotherapy, 40-50% of DLBCL patients remain incurable after first-line treatment with R-CHOP (rituximab, cyclophosphamide, adriamycin, vincristine, and prednisone). The disease is characterized by extreme clinical and prognostic heterogeneity, meaning two patients with apparently identical presentations can have dramatically different outcomes and treatment responses.
The precision medicine gap: Current standard prognostic tools, notably the International Prognostic Index (IPI), were built from clinical variables alone and offer only coarse risk stratification. They cannot account for the molecular diversity operating within and between tumors. To identify patients who will benefit from intensified or novel regimens, clinicians need biomarkers with higher specificity and sensitivity than IPI-based scoring can provide.
Systems biology as the solution: The authors argue that systems biology - the holistic integration of data from multiple biological layers - is better suited to capturing the complexity of DLBCL than any single-modality approach. This 2022 review systematically covers six omics layers applied to DLBCL prognostic biomarker discovery: genomics, transcriptomics, epigenetics, proteomics, metabonomics, and radiomics, along with emerging single-cell technologies. Biomarker samples are drawn from lymph node biopsies, spleen, bone marrow, peripheral blood, and cerebrospinal fluid (CSF), each with distinct clinical accessibility and information content.
The review was published in the International Journal of Biological Sciences and authored by researchers from Nanfang Hospital (Southern Medical University), Guangdong Medical University, Sun Yat-sen University, and Beijing Tongren Hospital. The paper is framed around a central dogma diagram connecting genomic alterations, through multi-omics layers, to single-cell resolution studies, illustrating how each layer contributes unique prognostic signal.
DLBCL subtyping has evolved through multiple generations of technology. The foundational 2000 work by Alizadeh et al. used DNA microarray gene expression profiling to divide DLBCL into germinal center B-cell (GCB) and activated B-cell (ABC) subtypes, with GCB patients showing significantly longer overall survival. Rosenwald et al. expanded this in 2002 with a third "unclassified" group (later combined with ABC into non-GCB). Hans et al. then proposed an IHC-based surrogate using CD10, BCL-6, and MUM1 expression as a cost-effective alternative to microarray. However, approximately 20% discordance exists between the Hans algorithm and genomic subtyping, requiring molecular confirmation when clinical findings conflict with IHC classification.
Next-generation sequencing subtypes: Schmitz et al. performed exome and transcriptome sequencing on 574 DLBCL biopsy samples combined with targeted amplicon resequencing of 372 genes, identifying four genetic subtypes: MCD (co-occurring MYD88-L265P and CD79B mutations), BN2 (BCL6 fusions and NOTCH2 mutations), N1 (NOTCH1 mutations), and EZB (EZH2 mutations and BCL2 translocations). BN2 and EZB subtypes had better survival than MCD and N1. Importantly, MCD is enriched in primary CNS DLBCL (37% of cases) and shows efficacy with BTK inhibitors. A limitation of this system is that 53.4% of patients could not be classified.
The LymphGen algorithm: Wright et al. developed the LymphGen probabilistic classifier, assigning DLBCL into seven genetic subtypes with distinct 5-year OS rates: MCD (40%), N1 (27%), A53 (63%), BN2 (67%), ST2 (84%), EZB-MYC+ (48%), and EZB-MYC- (82%). This subtyping system has been directly applied to guide personalized treatment: in a study by WL Zhao et al., ibrutinib was added to R-CHOP for MCD and BN2 subtypes, lenalidomide for N1 and NOS, decitabine for A53, and tucidinostat for EZB. The resulting R-CHOP+X genetic subtype-guided regimen achieved both higher complete response (CR) rates and longer progression-free survival (PFS) compared to standard R-CHOP.
Microenvironment-based subtyping: Kotlov et al. analyzed 4,580 DLBCL patients using 25 functional gene expression signatures (FGES) to reconstruct the lymphoma microenvironment (LME), identifying four subtypes: GC-like LME, Mesenchymal (MS-LME), Inflammatory (IN-LME), and Depleted (DP-LME). GC-like and MS-LME subtypes had approximately 80% 5-year survival, while IN-LME and DP-LME had poor prognosis, with DP-LME showing only approximately 60% 5-year patient survival. Population-based cohort analysis by Stuart et al. of 928 patients using targeted sequencing of 293 genes identified five molecular subtypes with 5-year OS ranging from 42.1% (MYD88 subtype, worst prognosis) to 64.9% (SOCS1/SGK1 subtype, best prognosis).
Beyond subtyping, genomic studies have identified individual driver mutations and circulating biomarkers with direct prognostic utility. MYC rearrangements occur in 5-10% of DLBCL, and approximately half of these also carry BCL2 rearrangements - the so-called "double-hit" configuration. Staiger et al. confirmed that dual expression of MYC and BCL2 proteins predicts poor prognosis independently. STAT3, located on chromosome 17, regulates cell growth through MYC upregulation; Huang et al. studied 185 R-CHOP-treated patients and demonstrated via IHC and siRNA assays that phosphotyrosine STAT3 (PY-STAT3) activation is strongly associated with poor survival, particularly in the ABC subtype.
SNP-based risk scoring: Ghesquieres et al. performed a meta-analysis of four studies and identified a two-SNP risk score capable of predicting event-free survival (EFS) with extraordinary statistical confidence (p = 1.78x10^-12), independent of treatment type, IPI, and cell-of-origin classification. The two trait loci were rs7712513 at 5q23.2 (near SNX2 and SNCAIP genes; HR 1.39, 95% CI 1.23-1.57 for EFS) and rs7765004 at 6q21 (near MARCKS and HDAC2 genes; HR 1.38, 95% CI 1.22-1.57 for EFS). Both loci were also independently predictive of OS, with HRs of 1.49 and 1.47, respectively.
Whole-exome sequencing findings: Pasqualucci et al. sequenced the whole-exome of 115 DLBCL samples and identified frequent mutations in histone/chromatin modifying genes (most commonly MLL2) and B2M (beta-2-microglobulin). Because B2M is required for T-cell immune recognition, its frequent mutation or deletion allows tumor cells to escape cytotoxic T-cell killing, establishing B2M deletion as a candidate prognostic indicator. Reddy et al. expanded this to 1,001 newly diagnosed patients using whole-exome and RNA sequencing, identifying 150 genetic drivers. They developed a genomic risk model yielding 5-year survival rates of approximately 60% (high-risk) vs. 90% (low-risk) in the complete remission group - outperforming the IPI, which separated the same groups at approximately 50% vs. 85%.
Circulating tumor DNA (ctDNA): ctDNA fragments released into the circulatory system by cancer cells contain tumor-specific genetic information and can be quantified via NGS without tissue biopsy. Sidaway et al. studied 217 DLBCL patients and found ctDNA was measurable in 98% prior to treatment. Dynamic ctDNA monitoring showed that early molecular response (EMR) and major molecular response (MMR) predicted 24-month EFS significantly better than imaging: EFS was 83% vs. 50% (P = 0.0015) for EMR, and 82% vs. 46% (P less than 0.001) for MMR. Roschewski et al. showed that metaphase ctDNA positivity was associated with a 5-year time-to-progression rate of 41.7% vs. 80.2% for ctDNA-negative patients, with sensitivity 47% and specificity 88%, performing better than PET/CT for outcome prediction.
Transcriptomic studies use RNA sequencing (RNA-seq), real-time quantitative PCR (qPCR), or microarray technology to measure transcript abundance across normal B lymphocytes, tumor cells, and distinct DLBCL subtypes. A key advantage over static genomics is the ability to capture gene expression at multiple treatment timepoints, providing dynamic prognostic information. Non-coding RNAs, including microRNAs (miRNAs) and long non-coding RNAs (lncRNAs), add an additional regulatory layer with distinct prognostic utility.
Double-hit signature (DHITsig): Ennishi et al. analyzed whole-exome sequencing, RNA-seq, and targeted resequencing from 157 GCB-DLBCL patients (including 25 high-grade B-cell lymphoma with MYC, BCL2, and/or BCL6 rearrangements, termed HGBL-DH/TH-BCL2). They constructed a 104-gene mRNA model called DHITsig. Kaplan-Meier survival analysis showed 5-year OS after R-CHOP treatment of 80% for DHITsig-negative patients vs. 60% for DHITsig-positive patients, establishing DHITsig as a strong transcriptomic prognostic tool within the GCB subtype.
Tumor microenvironment (TME) gene expression: Ciavarella et al. used the CIBERSORT algorithm to deconvolute gene expression data from 482 untreated DLBCL patients, identifying 45 TME genes. Expression of these 45 genes was then quantified using the NanoString technique in formalin-fixed, paraffin-embedded tissue. High vs. low TME gene expression yielded 30-month OS rates of approximately 90% vs. 62.5% (P = 0.00017) and PFS rates of approximately 75% vs. 60% (P = 0.0069). Combining COO with TME classification further sharpened stratification, with the combined model predicting 30-month OS below 25% in high-risk vs. above 75% in low-risk groups (P = 0.00022).
miRNA and lncRNA prognostic models: Sun et al. used miRNA PCR arrays to analyze 372 serum samples from 20 DLBCL patients across three treatment timepoints (diagnosis, remission, relapse). They identified a 4-miRNA prognostic model using miR-21, miR-130b, miR-155, and miR-28, which was significantly associated with worse PFS and OS in both training and validation cohorts by multivariate analysis. Zhuo et al. analyzed three GEO database cohorts (GSE31312, n = 426; GSE10846, n = 350; GSE4475, n = 129) and identified SubSigLnc-17, an lncRNA marker capable of distinguishing GCB from ABC subtypes with 92.5% sensitivity, while also providing prognostic predictive value. A key limitation of transcriptomics is its requirement for large sample sizes, making studies expensive, and its inapplicability when disease is driven by post-transcriptional rather than RNA-level alterations.
Proteomics identifies tumor-expressed proteins and their posttranslational modifications using two-dimensional liquid chromatography/tandem mass spectrometry (LC-MS/MS). Several surface and secreted proteins in DLBCL carry independent prognostic weight. PD-L1 expression on DLBCL tumor cells (but not on microenvironmental cells) is an independent prognostic factor for overall survival, as identified by Kiyasu et al., and serves as a therapeutic target for checkpoint inhibitors. FOXP1, an oncogenic transcription factor aberrantly expressed in DLBCL, drives down S1PR2 expression; low S1PR2, especially combined with high FOXP1, is an important predictor of poor prognosis since S1PR2 is required for germinal center B-cell homeostasis.
CD markers as biomarkers: In a cohort of 930 DLBCL patients, Niitsu et al. found 5-year OS rates of 55% vs. 65% for CD5+ (n = 102) vs. CD5- (n = 828) DLBCL, and 5-year PFS rates of 52% vs. 61%, respectively. Adding rituximab improved 4-year PFS in CD5+ patients from 47.4% to 62.5%, though OS improvement was not significant (57.8% vs. 63.5%), positioning CD5 as a predictive biomarker for rituximab response. Xu-Monette et al. demonstrated that CD37- DLBCL had significantly worse OS and PFS than CD37+ DLBCL after R-CHOP, establishing CD37 positivity as a favorable prognostic marker.
Serum free light chains (sFLC): Maurer et al., using the FREELITE assay in 219 patients, found that 32% had elevated pretreatment sFLC and 14% had abnormal kappa-to-lambda FLC ratios; both groups had poorer EFS and OS. Witzig et al. extended this with a 6-year follow-up in 276 untreated patients, finding elevated FLC (monoclonal group EFS: HR 3.56, 95% CI 1.88-6.76, P less than 0.0001; polyclonal group EFS: HR 2.56, 95% CI 1.50-4.38, P = 0.0006) to be a strong independent poor prognostic factor.
Epigenetic biomarkers: DNA methylation regulates gene expression and developmental processes. AICDA overexpression promotes intratumoral methylation heterogeneity; Teater et al. found that high AICDA expression - associated with elevated cytosine methylation heterogeneity - was a biomarker of poorer outcomes. N6-methyladenosine (m6A) is the most abundant internal co-transcriptional mRNA modification; Han et al. found that high expression of piRNA-30473 increased m6A levels and predicted poor prognosis, suggesting m6A methylation as a prognostic stratification tool. Histone deacetylation also suppresses CD20 expression in DLBCL, contributing to rituximab resistance, and Guan et al. showed that chidamide, a histone deacetylase inhibitor, synergizes with rituximab to restore tumor suppression.
Metabonomics analyzes all low-molecular-weight (50-1,500 Da) metabolites to identify pathophysiological changes in disease. In DLBCL, it complements genomics and proteomics by capturing downstream functional consequences of genetic and protein alterations. Monti et al. used consensus clustering of tumor cell gene expression to identify three DLBCL biological isoforms - oxidative phosphorylation (OxPhos), B-cell receptor/proliferation (BCR), and host response (HR) - with similar but not identical 5-year survival rates (OxPhos 53%, BCR/proliferation 60%, HR 54%; P = 0.53), suggesting that metabolic heterogeneity is present but does not directly map onto survival differences without additional stratifiers.
Glycolytic markers: Elevated glycolysis is a hallmark of tumor cells. Chiche et al. determined via unbiased analysis that GAPDH (glyceraldehyde-3-phosphate dehydrogenase) is the only glycolytic enzyme that independently predicts OS in R-CHOP-treated DLBCL. Multivariate analysis in 43 newly diagnosed patients showed high GAPDH expression was an independent predictor of improved OS (HR 0.603, P = 0.0371). Conversely, GOT2 (involved in alpha-ketoglutarate generation via the TCA cycle) showed the opposite: high GOT2 expression was significantly associated with shorter OS in 157 R-CHOP-treated patients (HR 2.28, P = 0.03756). A redox signature score incorporating markers such as VDUP1, MnSOD, ZnSOD, catalase, and thioredoxin reductase also stratified 5-year survival from 57% (low-risk quartiles) down to 37% (quartile 4, P less than 0.001 vs. quartile 1).
PET/CT radiomics: Radiomics, introduced by Dutch scholar Kumar in 2012, extracts quantitative features from CT, PET, and MRI data for disease prediction using machine learning and deep learning. PET/CT with 18F-FDG is now formally included in DLBCL staging and prognostic guidelines (since the 12th International Conference on Malignant Lymphoma). Interim PET/CT (iPET/CT) using the Deauville score combined with delta-SUVmax (the improved Deauville model) was evaluated in 593 R-CHOP-treated patients: iPET-negative patients had 3-year PFS of 80.2% and OS of 89.9% with CR rate 91.8%, while iPET-positive patients had only 3-year PFS of 12.5% and OS of 27.3% with CR rate 29.2%.
TMTV and radiomic texture features: Total metabolic tumor volume (TMTV) at baseline PET/CT is a strong survival predictor; elevated TMTV is significantly associated with poor PFS and OS, including in patients receiving lenalidomide maintenance. Lue et al. performed radiomic analysis on 83 DLBCL patients and identified RLNGLRLM (a gray-level run-length matrix texture feature from baseline FDG PET imaging) as an independent prognostic factor. High vs. low RLNGLRLM separated 5-year PFS at 37.2% vs. 91.7% and 5-year OS at 41.1% vs. 91.7%, a striking prognostic gap based on a single imaging texture feature. TMTV combined with spatial distribution parameters may further improve risk stratification at staging.
Bulk tissue multi-omics analyses average signals across heterogeneous cell populations, potentially obscuring biologically important subpopulations. Single-cell sequencing resolves this by characterizing genome, transcriptome, proteome, and epigenome at the individual cell level using high-throughput methods. For a highly heterogeneous malignancy like DLBCL - where even intratumoral heterogeneity at a single biopsy site is substantial - this resolution may reveal prognostic cell states invisible to bulk approaches.
scRNA-seq and CAR-T cell therapy response: Deng et al. studied 24 LBCL patients (16 DLBCL, 6 transformed follicular lymphoma, 2 PMBCL) who received autologous axicabtagene ciloleucel (axi-cel) anti-CD19 CAR-T cell therapy. After 3-month PET/CT follow-up, 50% of patients had progressive disease, 4% had partial remission, and 38% had complete remission. Whole transcriptome scRNA-seq of 137,326 residual cells revealed distinct T-cell profiles between responders and non-responders: CR patients showed significant enrichment of memory CD8+ T-cells (at approximately 3 times the level seen in PR/PD patients), while PR/PD patients showed enrichment of depleted CD8+ and CD4+ T-cells. Differentially expressed genes in CD8+ T-cells in CR vs. non-CR groups included BATF, ID2, IFNgamma, effector molecules (GZMA, GZMB, GNLY), and MHC class II molecules, which together serve as genetic markers of T-cell fitness and failure.
ctDNA correlation in CAR-T settings: In the same study, early molecular response (EMR) measured by ctDNA allele fold change on day 7 of CAR-T therapy was significantly correlated with clinical response (P = 0.008), providing an early liquid biopsy readout that parallels the scRNA-seq findings on T-cell fitness. This integration of single-cell transcriptomic profiling with ctDNA dynamics illustrates the potential of combining omics layers for real-time treatment monitoring.
Limitations and emerging platforms: scRNA-seq can only analyze the RNA profile of each cell once, and RNA capture efficiency is not fully stable, limiting the reproducibility and completeness of single-cell transcriptomes. Single-cell genomics, epigenomics, and proteomics have advanced rapidly but their application in DLBCL remains nascent. The challenge of low RNA capture efficiency, protein degradation, and noise in low-abundance analytes continues to hinder single-cell proteomics. Proteogenomics - combining proteomics and genomics - is emerging as a next-generation discipline that better reflects biological heterogeneity than mRNA abundance alone, since mRNA levels do not accurately predict protein abundance in cancer cells.
Biomarker sampling in DLBCL encompasses lymph node biopsy, spleen, bone marrow, peripheral blood, and cerebrospinal fluid (CSF). While lymphoma tissue biopsy remains the gold standard for molecular characterization, the inherent intratumoral heterogeneity of DLBCL means that a single-site biopsy may not accurately represent the full genetic landscape. Repeat biopsies during treatment are impractical, and genetic changes at relapse may not be captured. Liquid biopsy from peripheral blood - measuring ctDNA, circulating tumor cells (CTCs), and protein analytes - addresses these limitations by enabling non-invasive, multi-timepoint assessment.
Blood-based biomarkers beyond ctDNA: Rivas-Delgado et al. demonstrated that cfDNA from peripheral blood can substitute for tissue biopsy to assess tumor burden, mutational profile, and genetic classification with prognostic value in DLBCL. Bittenbring et al. found that Vitamin-D deficiency (VDD, serum levels less than or equal to 8 ng/mL) predicted poor prognosis: 3-year EFS was 59% vs. 79% (P adjusted HR 2.1, P = 0.008) and 3-year OS was 70% vs. 82% (HR 1.9, P = 0.040) for patients with vs. without VDD. VDD also reduced rituximab-mediated cytotoxicity (RMCC), and correcting VDD significantly increased RMCC (P less than 0.001). Absolute lymphocyte count (ALC) has also been proposed as a prognostic biomarker in the R(X)-CHOP era.
CSF biomarkers for CNS relapse: CNS is a major site of DLBCL relapse. IL-10/IL-6 ratio in CSF is elevated in primary CNS lymphoma (PCNSL) patients and correlates with poor prognosis. TIM-1 is associated with high IL-10 expression. Muñiz et al. identified elevated soluble CD19 (sCD19) in CSF as a predictor of CNS relapse. Elevated LDH in CSF or serum is consistently associated with increased CNS recurrence risk, including in adolescent DLBCL patients (ages 15-21, reported by Cairo et al.). Wang et al. found that elevated cfDNA concentrations in CSF correlate with high CNS-IPI scores, underscoring the role of CSF-cfDNA as a putative prognostic biomarker for detecting CNS tumor involvement. The six-factor CNS-IPI model (five IPI factors plus renal/adrenal involvement) is currently in clinical use for CNS risk stratification.
Future priorities: The authors identify several unmet needs: large-scale clinical studies validating individual biomarkers and signaling pathway interactions; evaluation of whether new molecular subtypes reliably predict responses to traditional and novel targeted drugs; development of single-cell technologies that reduce operational complexity, cost, and time while increasing sample throughput; and integration of multi-omics data through systems biology frameworks to build comprehensive, clinically deployable prognostic models. The convergence of single-cell histology, liquid biopsy, and AI-driven multi-omics integration is identified as the most promising near-term direction for DLBCL precision medicine.