Diffuse large B-cell lymphoma (DLBCL) is the most common type of non-Hodgkin lymphoma, accounting for roughly 30% of adult cases in western countries and an even higher proportion in developing nations. The standard treatment consists of CHOP (cyclophosphamide, doxorubicin, vincristine, and prednisone) or R-CHOP (CHOP combined with rituximab), which produces long-term disease-free survival in approximately 50% of patients. However, DLBCL is not a single disease. Gene expression profiling (GEP) has established that it comprises at least three molecular subtypes: germinal center B-cell like (GCB), activated B-cell like (ABC), and unclassified (UC).
The prognostic divide: These subtypes carry very different survival outcomes. GCB patients have a 5-year survival rate of approximately 59%, while ABC patients face a far worse 5-year survival rate of only 30%. The two subtypes also differ biologically. GCB is characterized by BCL2 translocations, REL loci amplifications, and immunoglobulin hypermutation, while ABC is defined by constitutive activation of the NF-kB pathway. This biological distinction has implications for treatment selection and drug development.
The gap between research and clinical practice: Despite its prognostic power, GEP has not been widely adopted in clinical practice because it requires measuring the expression of over 1,000 genes, which is expensive and technically demanding. Several immunohistochemistry (IHC)-based alternatives have been proposed (Hans algorithm, Choi algorithm, Visco-Young algorithm), but these have shown limited concordance with GEP. The Hans algorithm achieved only 71% concordance for GCB and 88% for Non-GCB subtyping. The Choi algorithm reached 83%, and the Visco-Young algorithm performed best at 92.6%. Additionally, IHC-based methods suffer from variability in staining quality, experimenter technique, and subjective interpretation of results.
This study aimed to bridge the gap by developing a machine learning model using only eight gene markers that could classify DLBCL patients into GCB and Non-GCB subtypes with high concordance to GEP, while remaining practical enough for clinical application.
Data sources: The authors obtained gene expression data from the Gene Expression Omnibus (GEO) database, all processed on the same platform (Affymetrix Human Genome U133 Plus 2.0 Array, GPL570). Expression values were normalized using three housekeeping genes (ACTB, GAPDH, and LDHA). The training set consisted of 414 patients from GSE10846, and validation used 855 patients from four additional datasets (GSE19426, GSE53786, GSE56315, and GSE31312). In total, 225 patients had been treated with CHOP and 778 with R-CHOP.
Gene screening process: Starting from 57 candidate genes previously associated with DLBCL prognosis or drug resistance, the authors applied two independent filtering methods. First, a Kruskal-Wallis H test with Nemenyi post-hoc testing identified 11 genes with significantly different expression across GCB, ABC, and UC subtypes. Second, a forward stepwise Fisher discriminant analysis selected 20 genes capable of separating the three subtypes. The intersection of these two approaches yielded eight common genes: MYBL1, LMO2, BCL6, MME, IRF4, NFKBIZ, PDE4B, and SLA.
Choosing the machine learning method: Seven machine learning algorithms were compared using fivefold cross-validation: decision tree, random forest, support vector machine (SVM), Fisher discriminant analysis, nearest neighbor, bagging, and AdaBoost. The comparison was run both on the full 57-gene set and on the reduced 8-gene set. SVM achieved the minimum classification error rate within the 8-gene model, and its error rate was actually smaller than most methods using all 57 genes. This made SVM the clear choice for the final classification model, implemented using the R package "e1071."
Cutoff establishment: Rather than using mean or median expression values (which would be inappropriate for the non-Gaussian distribution of gene expression data), the authors used receiver-operating characteristic (ROC) curves and calculated the Youden index for each gene. This identified the expression value that maximized the combined sensitivity and specificity for classifying a sample as GCB versus Non-GCB according to GEP analysis.
GCB-associated genes (Group 1): Correspondence analysis confirmed that four genes cluster with the GCB subtype. MYBL1, a member of the Myb oncogene family of transcription factors, is specifically induced in proliferating centroblasts and located in the chromosome 8q22 region involved in recurrent lymphoma translocations. LMO2, the strongest single predictor of overall survival in DLBCL, is part of the germinal center expression profile and is associated with better survival in CHOP/R-CHOP-treated patients. BCL6 is frequently translocated and hypermutated in DLBCL, promoting GCB cell proliferation in response to T-cell-dependent antigens. MME (also known as CD10 or neutral endopeptidase) cleaves cell surface signal peptides affecting proliferation, differentiation, and migration, and its reduced expression correlates with less favorable outcomes.
Non-GCB-associated genes (Group 2): The remaining four genes cluster with the ABC and UC subtypes. IRF4 sits at the center of B-cell terminal differentiation and ABC-DLBCL transcriptional programs, driven by a positive feedback loop involving NF-kB, CARD11, and B-cell receptor signaling. PDE4B deactivates cAMP and abolishes its inhibitory effects in B lymphocytes, and its overexpression is associated with fatal or refractory tumors and resistance to cAMP-induced apoptosis via the PI3K/AKT pathway. NFKBIZ encodes IkB-zeta, which regulates the constitutively activated NF-kB pathway that is a hallmark of ABC-DLBCL. SLA encodes an adaptor protein that negatively regulates tyrosine kinase signaling in lymphocytes and may contribute to the worse chemotherapy response seen in ABC patients.
Expression patterns and cutoffs: All eight genes showed significantly different expression between GCB and Non-GCB subtypes (P = 0.000 for all comparisons). The GCB-associated genes had significantly higher expression in GCB cases. For example, MYBL1 had a cutoff of 9.3 (specificity 90.7%, sensitivity 72.7%), LMO2 a cutoff of 12.3 (specificity 72.1%, sensitivity 75.8%), BCL6 a cutoff of 13 (specificity 79.2%, sensitivity 79.2%), and MME a cutoff of 8.5 (specificity 73.8%, sensitivity 87.4%). Conversely, the Non-GCB genes showed higher expression in ABC cases, with IRF4 at cutoff 10.7 (specificity 69.9%, sensitivity 74.9%), NFKBIZ at 10.08 (specificity 63.9%, sensitivity 76.6%), PDE4B at 11.5 (specificity 68.3%, sensitivity 79.7%), and SLA at 10.5 (specificity 78.7%, sensitivity 70.1%).
Univariate analysis confirmed all eight genes had significant prognostic power for overall survival (P < 0.05). In multivariate Cox regression, three genes reached independent significance: LMO2 (P = 0.009), BCL6 (P = 0.038), and IRF4 (P = 0.023). The hazard ratios were consistently negative for GCB-associated genes (indicating favorable prognosis) and positive for Non-GCB-associated genes (indicating poor prognosis).
In the training cohort of 414 patients from GSE10846, GEP classified patients as GCB (183, 44.2%), ABC (167, 40.6%), or unclassifiable (64, 15.2%). The eight-gene SVM model correctly classified 94.0% of patients as either GCB or Non-GCB according to GEP analysis. According to the model, 170 patients (41.1%) had a GCB phenotype and 244 (58.9%) had a Non-GCB phenotype. The 64 unclassifiable cases were assigned to GCB (9 cases) or Non-GCB (55 cases) by the model, providing a classification for patients that GEP could not categorize.
Concordance with GEP: For the 350 patients classified by GEP as either GCB or ABC (excluding unclassified cases), the model achieved 99.4% concordance for GCB (only 1 mismatch out of 170 GCB patients) and 99.2% concordance for ABC (14 mismatches out of 180 ABC patients). This level of agreement is substantially higher than any previously reported IHC-based algorithm.
Validation cohorts: The model was validated in two independent cohorts. Test1 included 380 patients from GSE19426, GSE53786, and GSE56315, where the model achieved 91.0% concordance with GEP. Test2 included 475 patients from GSE31312, where concordance reached 94.4%. Both validation results confirmed that the model generalizes well to independent patient populations.
Why eight genes, not fewer? The authors systematically tested whether fewer genes could achieve the same accuracy. A three-gene model (LMO2, BCL6, IRF4, the three multivariate-significant genes) had high concordance for GCB (93.4% training, 97.8% testing) but poor concordance for ABC (79.3% training, 63.9% testing). Models with 4, 5, 6, and 7 genes were each incrementally better. Only the full eight-gene model achieved balanced, high concordance for both subtypes across all datasets, confirming it as the minimum effective combination.
The authors directly compared their SVM model's classification error rates against established IHC-based algorithms on the same patient cohorts. In the GSE53786 dataset, the SVM model achieved an error rate of only 2%, compared to 9% for the Hans algorithm and 14% for the Choi algorithm. In GSE31312, the SVM model had a 5.6% error rate, compared to 7.45% for the 3-marker Visco-Young algorithm and 7.15% for the 4-marker Visco-Young algorithm.
Diagnostic power: ROC curve analysis further quantified the model's discriminatory accuracy. The area under the curve (AUC) was 0.989 (P = 0.000) in the training data and 0.978 (P = 0.000) in the validation data. Individual gene markers also showed strong diagnostic power. The GCB-associated genes (MYBL1, LMO2, BCL6, MME) each had AUC values above 0.8 for identifying GCB patients, while the Non-GCB genes (IRF4, PDE4B, NFKBIZ, SLA) had AUC values above 0.7 for identifying ABC patients.
These results indicate that the SVM model with eight gene markers provides a more accurate molecular subtype classification than any existing IHC-based approach. The model achieved high sensitivity and specificity for both subtypes: 87% sensitivity and 90% specificity for ABC, and 92% sensitivity and 93% specificity for GCB.
The central question for any molecular classifier is whether it predicts patient outcomes as effectively as the gold-standard method. In the training cohort (414 patients, median follow-up 2.38 years), the 5-year overall survival (OS) was 57.71%. When stratified by GEP, the 5-year OS was 71.48% for GCB versus 46.87% for Non-GCB (HR 0.451, 95% CI 0.331-0.614, P < 0.0001). The SVM model produced nearly identical stratification: 72.94% for GCB versus 46.29% for Non-GCB (HR 0.445, 95% CI 0.326-0.606, P < 0.0001).
Validation in independent cohorts: In 119 patients from GSE53786 treated with CHOP or R-CHOP, the model successfully stratified patients into groups with significantly different OS (HR 0.434, 95% CI 0.239-0.789, P = 0.006). In 475 R-CHOP-treated patients from GSE31312, similar stratification was observed for both OS (HR 0.597, 95% CI 0.441-0.807, P = 0.0008) and progression-free survival (PFS; HR 0.596, 95% CI 0.441-0.806, P = 0.0007). These validation results closely matched the prognostic power of GEP itself (PFS by GEP: HR 0.612, 95% CI 0.445-0.842, P = 0.0025).
Clinical characteristics: The model's subtype assignments correlated with known clinical features. Non-GCB patients were significantly older (median 64.5 vs. 60 years), had more advanced stage disease (57.8% vs. 48.3% stage III-IV), and higher IPI scores (33.7% vs. 22.9% with IPI 3-5). Multivariate analysis confirmed that IPI score 3-5 (HR 2.437, P = 0.007), advanced stage (HR 1.159, P = 0.027), high LDH (HR 1.135, P = 0.000), poor performance status (HR 1.336, P = 0.016), and older age (HR 1.026, P = 0.004) were independent adverse prognostic factors.
The International Prognostic Index (IPI) is the most widely used clinical tool for predicting DLBCL outcomes, incorporating age, stage, LDH level, performance status, and number of extranodal sites. A critical question is whether the eight-gene SVM model provides additional prognostic information beyond what IPI already captures. To test this, the authors stratified patients by IPI score (low: 0-2, high: 3-5) and then further subdivided each IPI group into GCB and Non-GCB using the model.
Results in the training cohort: Within the low-IPI group, the model separated patients into GCB (5-year OS 80.5%) and Non-GCB subgroups with significantly different survival (HR 0.436, 95% CI 0.274-0.694, P = 0.0005). Within the high-IPI group, the separation was equally informative, identifying a Non-GCB subgroup with a 5-year OS of only 14.6% (HR 0.426, 95% CI 0.170-1.068, P = 0.0068). This means combining IPI with the SVM model identified the best-prognosis group (low IPI + GCB, 80.5% 5-year OS) and the worst-prognosis group (high IPI + Non-GCB, 14.6% 5-year OS) with striking separation.
Validation in GSE31312: The same pattern held in the validation cohort of 475 R-CHOP-treated patients. In the low-IPI group, the GCB subtype predicted 70.2% 5-year OS (HR 0.648, P = 0.033). In the high-IPI group, Non-GCB patients had only 41.4% 5-year OS (HR 0.542, P = 0.017). These consistent results across training and validation cohorts confirm that the model is independent of IPI and adds meaningful prognostic stratification that IPI alone cannot provide.
Retrospective design: All data in this study came from publicly available GEO datasets, meaning the analysis is entirely retrospective. The patients were treated at different institutions with varying protocols, and while this heterogeneity strengthens the generalizability argument, it also introduces confounding variables that a prospective trial would control. No prospective validation was performed, and the model has not been tested in a real-time clinical decision-making setting.
Platform dependency: The model was built exclusively on Affymetrix U133 Plus 2.0 Array data processed with MAS 5.0 normalization. It is unclear how the model would perform with data from other gene expression platforms (such as NanoString nCounter, RNA-seq, or different microarray platforms). Cross-platform validation would be necessary before the model could be considered platform-agnostic. Additionally, the cutoff values were derived from this specific platform and may not transfer directly to other measurement technologies.
Binary classification: The model classifies patients into GCB and Non-GCB, collapsing the ABC and UC subtypes into a single Non-GCB category. While this is clinically useful (the key prognostic distinction is GCB vs. Non-GCB), it discards potentially relevant information about the unclassified subtype. The 64 UC cases in the training set were assigned to GCB (9) or Non-GCB (55) by the model, but whether these assignments carry true prognostic meaning for UC patients specifically could not be statistically confirmed due to the small number assigned to GCB. Gender data was not available for the validation set, preventing full analysis of gender as a prognostic variable.
Clinical translation: For this model to reach clinical practice, the eight-gene expression panel would need to be implemented on a clinically accessible platform, validated prospectively across multiple institutions, and compared head-to-head with newer molecular methods such as the NanoString Lymph2Cx assay. The authors note that the model could be particularly valuable for guiding treatment decisions in the era of targeted therapies, where ABC-specific agents like lenalidomide and ibrutinib are being investigated. The combination of IPI scoring with eight-gene molecular subtyping could enable more precise risk stratification and more individualized treatment selection for DLBCL patients.