Mantle cell lymphoma (MCL) and peripheral T-cell lymphoma (PTCL) represent two relatively rare but clinically aggressive subtypes of adult-onset non-Hodgkin lymphoma (NHL). MCL arises from mature B cells in the mantle zone of lymphoid follicles and accounts for roughly 6% of all NHL cases. It is characterized by the t(11;14) translocation, which overexpresses cyclin D1 and drives aberrant cell-cycle progression. Standard treatment combines rituximab with intensive chemotherapy regimens, typically incorporating high-dose cytarabine (HiDAC), followed by autologous stem cell transplantation (auto-SCT) in younger, fit patients. Despite these aggressive approaches, relapse is nearly universal for most patients, and median overall survival ranges from 3 to 5 years in most series, though a minority of patients with indolent or low-risk disease can be managed conservatively.
PTCL as a clinical entity: Peripheral T-cell lymphomas encompass a heterogeneous group of post-thymic T-cell neoplasms with notably poor prognosis. The most common subtypes include PTCL-not otherwise specified (PTCL-NOS), angioimmunoblastic T-cell lymphoma (AITL), and anaplastic large cell lymphoma (ALCL). Across the group, 5-year overall survival rates for most PTCL subtypes treated with CHOP-based chemotherapy hover between 20% and 40%, with the exception of ALK-positive ALCL, which carries a more favorable prognosis. Most patients receive front-line chemotherapy followed by consolidative auto-SCT, but response rates and durability are considerably lower than in aggressive B-cell lymphomas.
The prognostic gap: Despite progress in characterizing the genomic drivers and molecular heterogeneity of both MCL and PTCL, validated molecular prognostic tools have not been widely adopted in routine clinical practice. Clinical prognostic indices, such as the MIPI (Mantle Cell Lymphoma International Prognostic Index) for MCL and the PIT (Prognostic Index for T-cell lymphoma) for PTCL, rely solely on clinical variables and cannot stratify patients based on the tumor's underlying biology. The low incidence of these lymphomas limits the accumulation of large discovery cohorts, creating a bottleneck for the development and validation of robust molecular prognostic signatures.
This report from Mosquera-Orgueira and colleagues at the University Hospital of Santiago de Compostela, presented at the EHA2023 Hybrid Congress and published in HemaSphere, addresses this gap directly by asking whether gene expression-based machine learning models originally developed for other lymphoma subtypes can be repurposed to stratify risk in MCL and PTCL patients.
The study builds directly on two prior machine learning prognostic frameworks developed at the same institution. The first, called LymForest, was originally constructed for diffuse large B-cell lymphoma (DLBCL), the most common aggressive NHL subtype. LymForest combines a curated gene expression signature with a random forest survival model to derive personalized risk scores for DLBCL patients. The authors had previously reported that LymForest achieves strong discriminative performance in DLBCL cohorts, providing prognostic information beyond the standard International Prognostic Index (IPI). This prior publication (Mosquera et al., 2022) established the methodological foundation for the current study.
Iacobus-FL: The second signature, Iacobus-FL, was developed for follicular lymphoma (FL), an indolent B-cell NHL subtype. FL is biologically distinct from DLBCL, but shares some transcriptional features with other germinal center-derived B-cell lymphomas. Iacobus-FL uses a different gene panel compared to LymForest but is interpreted using the same random forest survival modeling framework. Together, these two signatures cover the most common indolent and aggressive B-cell NHL subtypes.
Scientific rationale for cross-subtype testing: The hypothesis underlying this study is that gene expression patterns relevant to lymphoma biology and prognosis may not be entirely subtype-specific. Many of the genes driving tumor behavior, such as those governing cell proliferation, immune evasion, stromal interaction, and apoptosis resistance, are dysregulated across multiple lymphoma subtypes, even if the precise combination of aberrations differs. If LymForest and Iacobus-FL capture broadly relevant transcriptional programs, they may retain prognostic signal when applied to different lymphoma entities. Testing this cross-subtype transferability is both scientifically informative and practically valuable: it could reduce the need to develop entirely new signatures for every rare lymphoma subtype.
The decision to test these B-cell signatures in PTCL is particularly noteworthy given the biological differences between B-cell and T-cell lymphomas. T-cell lymphomas originate from a completely different cell lineage and carry distinct gene expression profiles. The finding that a B-cell-derived signature retains any prognostic value in PTCL would suggest that some transcriptional prognostic programs are conserved across lymphoid malignancies.
The analytical approach follows the framework established in the authors' 2022 DLBCL paper. A random forest model of survival was constructed according to the same prior specifications, meaning the gene expression features and model architecture were not re-optimized for MCL or PTCL. This is a critical methodological point: the model was applied in a transfer learning context, with gene selection and hyperparameters fixed to those determined in the original lymphoma cohorts. Any prognostic performance observed in MCL and PTCL therefore represents genuine out-of-domain generalization rather than in-sample optimization.
Out-of-bag estimation: To adjust for optimism in model performance estimates, the authors used out-of-bag (OOB) prediction. In ensemble methods like random forests, each tree in the forest is trained on a bootstrap sample of the data, leaving out approximately one-third of observations. These left-out cases are predicted by the trees that did not include them in training, providing an unbiased estimate of model performance without requiring a separate hold-out test set. OOB estimation is particularly valuable when sample sizes are small, as it efficiently uses all available data while still generating optimism-corrected performance metrics.
Survival predictions: Model outputs were based on the cumulative hazard function and the survival function derived from the random forest ensemble. For each patient, the model produces a continuous risk score derived from the ensemble of survival trees. Patients were then divided into three equal-sized risk subgroups (tertiles) based on these scores, enabling Kaplan-Meier survival curve analysis and visual assessment of risk discrimination. The primary performance metric was Harrell's concordance index (C-index), which measures the probability that the model correctly ranks pairs of patients by their true survival time (analogous to AUC for survival data, where C = 0.5 indicates no discrimination and C = 1.0 indicates perfect discrimination).
External transcriptomic datasets: Rather than generating new gene expression data, the authors leveraged publicly available transcriptomic datasets from the Gene Expression Omnibus (GEO). The MCL cohort was drawn from Stott et al. (2017), and the PTCL cohort from Iqbal et al. (2014). Both are established reference datasets in the lymphoma field, providing independent external validation cohorts for the gene expression signatures. Using pre-existing GEO data is standard practice for rare lymphoma subtypes, where prospective collection of sufficiently large cohorts within a single institution is not feasible.
The primary result in the MCL cohort is described by the authors as "perfect reproducibility" of the LymForest signature. The model achieved a C-index of 0.721 for overall survival in MCL patients, indicating that in roughly 72% of patient pairs, the model correctly assigned higher risk scores to the patient who experienced the event first. A C-index of 0.721 is considered a strong result in lymphoma prognostication, particularly for a model applied out-of-domain without any MCL-specific retraining. For context, clinical prognostic indices such as the MIPI typically achieve C-index values in the range of 0.60 to 0.68 in independent validation cohorts, suggesting that the gene expression-based LymForest model provides meaningfully better discrimination than standard clinical variables alone.
Kaplan-Meier stratification: Beyond the overall C-index, the model successfully stratified MCL patients into three prognostically distinct groups of equal size. The Kaplan-Meier survival curves for these three tertiles showed clear separation, with statistically significant and clinically meaningful differences in overall survival between the low, intermediate, and high-risk groups. This three-way stratification is clinically important because it identifies not just a binary low/high risk split but also an intermediate group that may represent candidates for intensified therapy or clinical trial enrollment.
Why MCL biology may explain cross-subtype applicability: MCL, although a B-cell lymphoma like DLBCL, carries distinct molecular features, including cyclin D1 overexpression, frequent TP53 mutations, and complex karyotypes in aggressive cases. However, many of the transcriptional programs captured by LymForest, such as those governing tumor microenvironment, cell cycle regulation, and apoptosis pathway activity, are likely relevant across multiple B-cell malignancies. The strong cross-subtype performance of LymForest in MCL suggests that the model has captured robust biological prognostic signals rather than DLBCL-specific artifacts.
The result also has practical implications. Gene expression profiling of MCL for prognostic purposes has been investigated previously, with signatures such as the "proliferation signature" (reflecting tumor cell cycling rate) identified as among the strongest prognostic factors in MCL. The LymForest signature may be capturing overlapping biology, including proliferation-related transcripts, given that DLBCL and MCL both show strong prognostic associations with proliferative gene expression programs.
The PTCL results represent the more scientifically unexpected finding of the study. Testing a B-cell-derived gene expression signature in a T-cell lymphoma entity requires that at least some of the transcriptional features captured by LymForest and Iacobus-FL are expressed and prognostically relevant in T-cell neoplasms as well. The combined application of both LymForest and Iacobus-FL signatures achieved a C-index of 0.626 for overall survival in the PTCL cohort. While this is lower than the MCL result, a C-index of 0.626 still indicates meaningful prognostic discrimination in a biologically distinct and heterogeneous entity where even established clinical indices like the PIT perform modestly (PIT C-index typically 0.58 to 0.65 in validation cohorts).
Three-group risk stratification in PTCL: As in the MCL analysis, patients were divided into three equal-sized risk tertiles, and the resulting Kaplan-Meier curves showed divergent overall survival between groups. The term "divergent" used by the authors implies that the three survival curves do not overlap, confirming that the risk score meaningfully separates PTCL patients with meaningfully different prognoses. This result is described as confirming "reasonable predictability" in PTCL, a measured characterization that appropriately reflects both the statistical significance of the finding and the acknowledged limitations of applying B-cell-trained models to T-cell disease.
Why both signatures were combined for PTCL: The use of the combined LymForest and Iacobus-FL signature for PTCL, rather than LymForest alone as in MCL, suggests that the combination captures a broader range of relevant biological processes. PTCL-NOS is itself a heterogeneous category encompassing tumors with gene expression profiles spanning multiple transcriptional programs, including some that overlap with germinal center biology (relevant to FL, the substrate for Iacobus-FL). The combination may improve coverage of this molecular diversity by drawing on signals from both signatures.
Biological interpretation of cross-lineage applicability: Transcriptomic prognostic programs centered on immune microenvironment composition, stromal remodeling, and tumor proliferation rate are active areas of research across B-cell and T-cell lymphomas. Prognostic gene expression models that capture high tumor microenvironment activity, T-regulatory cell infiltration, or macrophage polarization states would be expected to retain prognostic relevance across lineages, since these microenvironmental factors influence patient outcomes independent of the specific malignant cell type. This cross-lineage biology likely underlies the retained prognostic signal observed in PTCL.
The principal clinical implication of these findings is the potential to extend validated molecular prognostic tools to rare lymphoma subtypes without requiring de novo signature development for each entity. In a rare disease context, generating a sufficiently large discovery cohort to develop and validate a new gene expression signature from scratch requires multi-institutional collaboration and years of sample collection. The cross-subtype applicability demonstrated here offers a practical shortcut: clinicians and researchers could apply existing validated tools such as LymForest to MCL and PTCL patients immediately, provided the same gene expression platform (e.g., NanoString nCounter, RNA sequencing) is available.
Personalized risk scoring in MCL: Current MCL prognostication relies primarily on the MIPI, which integrates age, ECOG performance status, LDH, and leukocyte count. While biologic factors such as Ki-67 proliferation index and TP53 mutation status are increasingly incorporated, standardized gene expression-based risk scores are not yet part of routine practice. A validated tool like LymForest achieving C-index 0.721 would represent a meaningful improvement over clinical-variable-only indices and could inform decisions about transplant candidacy, maintenance therapy with BTK inhibitors, or enrollment in clinical trials testing novel agents such as venetoclax or CAR-T cell therapies in MCL.
PTCL management implications: In PTCL, the heterogeneity of the disease means that refining risk stratification has direct implications for treatment intensity. Patients identified as high-risk by gene expression scoring could be prioritized for upfront auto-SCT consolidation, allogeneic stem cell transplantation in remission, or enrollment in trials testing novel agents such as brentuximab vedotin combinations, PI3K inhibitors, or immune checkpoint blockade. Low-risk patients might be spared the morbidity of aggressive consolidation without compromising outcomes, a critical goal given the toxicity burden of current PTCL treatment regimens.
The ability to use a single gene expression platform to risk-stratify patients across DLBCL, follicular lymphoma, MCL, and PTCL would also have significant practical value for diagnostic laboratories, reducing the need for disease-specific assays and potentially lowering the cost and turnaround time for molecular risk stratification in lymphoma care.
Abstract format and data availability: The most significant limitation of this report is that it was presented as a conference abstract at the EHA2023 Hybrid Congress rather than as a full peer-reviewed research article. As a result, the sample sizes of the MCL and PTCL cohorts are not explicitly stated, the specific genes included in the LymForest and Iacobus-FL panels are not described, and the full analytical details, including confidence intervals around the C-index, log-rank test p-values, and covariate adjustments, are not available. Evaluating the robustness of the reported results requires access to the complete manuscript data.
Small cohort sizes in rare subtypes: The external GEO datasets used for validation, Stott et al. (2017) for MCL and Iqbal et al. (2014) for PTCL, are both relatively small by modern standards. The Stott MCL dataset contains fewer than 100 samples and the Iqbal PTCL dataset contains approximately 200 samples with heterogeneous subtypes. With such small cohorts, the C-index estimates have wide confidence intervals, and dividing patients into three equal tertiles may result in groups of only 20 to 30 patients each, limiting the statistical power to detect significant Kaplan-Meier differences. The clinical composition of these cohorts, including treatment regimens used and whether patients underwent auto-SCT, is also not described in the abstract.
Prospective validation absent: The analysis is entirely retrospective and relies on archival gene expression data collected in the 2010s. Treatment standards for both MCL and PTCL have evolved substantially since these datasets were assembled. Newer MCL patients may receive BTK inhibitors (ibrutinib, acalabrutinib) as part of front-line therapy or at relapse, altering the prognostic landscape considerably. The relevance of the LymForest signature to patients treated on contemporary protocols requires prospective validation in current clinical trial cohorts.
Platform and standardization considerations: Gene expression signatures developed on one measurement platform (e.g., Affymetrix microarray) may not translate directly to other platforms (e.g., RNA-seq, NanoString) without analytical adjustment. The GEO datasets used here were likely generated on microarray platforms, while clinical implementation would require platform-specific normalization and possibly re-training of signature thresholds. These practical translation steps are not addressed in the abstract.
Prospective validation in current treatment contexts: The authors explicitly envision evaluating these signatures "in the context of innovative therapeutic schemes," acknowledging that retrospective validation in archival cohorts is a first but insufficient step. For MCL, prospective validation should occur in cohorts that include patients treated with BTK inhibitor-containing regimens, venetoclax combinations, and CAR-T cell therapies in the relapsed setting. For PTCL, relevant contemporary contexts include brentuximab vedotin plus CHP (BV-CHP) for CD30-positive subtypes, and novel approaches such as anti-CD47 agents, histone deacetylase (HDAC) inhibitors, and immune checkpoint blockade for CD30-negative entities.
Standardization for clinical implementation: Moving from research-grade GEO microarray data to a standardized clinical assay requires several steps. The gene panels for LymForest and Iacobus-FL would need to be translated onto a clinical-grade platform such as NanoString nCounter or a targeted RNA-seq panel. Analytical thresholds for risk tertile assignment would need to be recalibrated against contemporary patient populations. The assay would require analytical validation for reproducibility, sensitivity, and specificity before regulatory submission and clinical use.
Expanding to additional rare subtypes: The successful application of these signatures in MCL and PTCL opens the question of whether they retain prognostic value in other rare NHL subtypes, such as extranodal NK/T-cell lymphoma, primary mediastinal B-cell lymphoma (PMBCL), or transformed marginal zone lymphoma. A systematic cross-subtype evaluation using available GEO datasets could map the boundaries of signature transferability across the NHL spectrum, potentially establishing a universal lymphoma prognostic framework applicable with a single gene expression test.
Multimodal integration: Gene expression-based risk scores provide molecular prognostic information complementary to clinical variables, imaging features (PET/CT-derived total metabolic tumor volume, TMTV), and genetic mutation profiles (TP53, CDKN2A in MCL; TET2, RHOA, IDH2 in PTCL). Future work integrating LymForest-derived risk scores with baseline PET/CT imaging and targeted sequencing data into multimodal machine learning models could achieve substantially higher prognostic accuracy than any single data modality alone, moving toward truly individualized treatment selection for aggressive NHL.