Determining clinical course of diffuse large B-cell lymphoma using targeted transcriptome and machine learning algorithms

Diffuse large B-cell lymphoma (DLBCL) is the most common subtype of lymphoma, exhibiting significant heterogeneity in its clinical course and patient outcomes. While over 60% of DLBCL patients achieve a cure with rituximab, cyclophosphamide, doxorubicin, vincristine, and prednisone (R-CHOP) treatment, the disease's heterogeneity necessitates more precise treatment strategies. Several approaches have attempted to subclassify DLBCL based on biological characteristics. Early methods used microarray-based expression profiling to categorize DLBCL into germinal center B-cell-like (GCB) and activated B-cell-like (ABC) subtypes based on cell of origin (COO). However, this approach left 15% unclassified and lacked sufficient predictive power for overall survival (OS) and progression-free survival (PFS). Subsequent refinements, such as the GenClass and LymphGen algorithms, incorporated genetic abnormalities but still classified only a fraction of cases and failed to reliably predict clinical outcomes. Other approaches like mutation profiling and chromosomal structural analysis (gains and losses) also proved insufficient in clinically relevant prediction. The co-occurrence of MYC rearrangements with BCL2, BCL6, or both, indicates a particularly aggressive DLBCL that often does not respond to R-CHOP. While existing subclassification strategies can distinguish biologically distinct DLBCL subgroups, they fall short in accurately predicting patient survival and are hindered by the requirement for whole-exome sequencing, making clinical implementation challenging. The researchers hypothesized that RNA profiling, reflecting changes in gene expression resulting from chromosomal structural alterations and mutations, would provide a more practical and clinically applicable approach to DLBCL subclassification.

The literature review section extensively examines existing methods for DLBCL subclassification, highlighting their limitations. It discusses the limitations of microarray-based expression profiling, which resulted in the identification of GCB and ABC subtypes, but failed to classify 15% of cases and did not accurately predict patient outcomes. The review also highlights the shortcomings of the GenClass and LymphGen algorithms, which were based on genetic abnormalities and still lacked sufficient predictive power. Furthermore, the literature review covers methods that incorporated mutation profiling and chromosomal structural analysis but also fell short of providing robust clinical predictions. Finally, the review points to the difficulty of clinical implementation of these methods due to their complexity and the need for whole-exome sequencing. This sets the stage for the researchers' proposed novel approach based on targeted RNA sequencing and machine learning.

This study developed a DLBCL classification strategy using targeted RNA sequencing combined with machine learning algorithms to predict clinical outcomes. The researchers analyzed data from 379 patients with de novo DLBCL and 247 patients with extranodal DLBCL, all treated with R-CHOP at 22 medical centers. RNA was extracted from formalin-fixed paraffin-embedded (FFPE) tissue, and a targeted RNA sequencing panel capturing 1408 cancer-associated genes was utilized. Data analysis used a novel machine learning approach that first grouped patients based on survival data, using a modified Bayesian approach that also handles censored survival data. This involved creating a survival model using a generalized naïve Bayesian classifier to predict the survival of censored patients to overcome the limitations of standard machine learning techniques in handling this type of data. The researchers addressed the challenge of numerical underflow in high-dimensional data by using a geometric mean to calculate the likelihood products. A 12-step cross-validation strategy was implemented to select a set of genes that effectively predicted survival subgroups and prevent overfitting. The algorithm divided patients into four survival groups (LL, LS, SL, SS) based on their overall survival. The selected genes were then validated using an independent set of 247 extranodal DLBCL samples. Multivariate Cox proportional hazard regression was used to assess the independent prognostic value of the survival classification compared with other clinical parameters, including COO (GCB/ABC), IPI, TP53 mutations, MYD88 and CD79B mutations, and MYC and IRF4 expression levels.

The study successfully developed a machine learning model that stratified DLBCL patients into four survival subgroups based on their overall survival, effectively predicting survival characteristics in both the training and validation cohorts. The model utilized a selected set of 180 genes to predict the four survival subgroups. Validation using an independent set of 247 extranodal DLBCL patients confirmed the model's robustness. Multivariate analysis revealed that the new survival classification, combined with the International Prognostic Index (IPI), was a significant predictor of survival outcomes, with TP53 mutations remaining an independent prognostic biomarker. The survival groups correlated with cell of origin (COO), TP53 mutations, MYC expression, and IRF4 expression. Notably, despite similar overall survival, the LS and SL groups exhibited significantly different biological characteristics, demonstrating the model's ability to capture subtle but clinically important differences. This emphasizes the limitations of using single biomarkers for predicting clinical behavior in DLBCL. MYD88 mutations were associated with better survival. The model accurately predicted the overall survival and progression-free survival for the four identified subgroups. Extranodal DLBCL was associated with shorter survival. Combining datasets of nodal and extranodal DLBCL patients, two-thirds were used to train and one-third to test the model, and the results remained substantially similar.

This study offers a significant advancement in DLBCL prognosis prediction by integrating targeted transcriptome data with a robust machine-learning approach. The novel method circumvents the limitations of previous subclassification strategies by directly using survival outcomes to identify predictive biomarkers. The finding that only TP53 mutations remain an independent prognostic factor after accounting for the survival subgroups highlights the comprehensive nature of this novel approach. The identification of distinct gene sets associated with each survival group paves the way for targeted therapy development. The results emphasize the inherent heterogeneity of DLBCL and the limitations of relying on single biomarkers for accurate prognosis. The ability to reliably predict patient survival based on readily accessible RNA sequencing data opens up possibilities for personalized treatment strategies. Identifying patients who will not respond to standard R-CHOP therapy enables earlier intervention with alternative treatments or clinical trials.

This research presents a novel approach for classifying DLBCL patients into four distinct survival subgroups using targeted RNA sequencing and machine learning. This classification system is robust, validated in independent cohorts, and offers valuable prognostic information beyond existing methods. This approach has clinical utility for identifying patients who may not respond to standard R-CHOP therapy, allowing for timely adjustments to treatment strategies or enrollment in targeted clinical trials. Future research could focus on exploring the biological mechanisms underlying the identified gene sets and developing targeted therapies based on these findings.

While this study used a large dataset and rigorous validation techniques, certain limitations exist. The study population was primarily composed of patients treated with R-CHOP; therefore, the generalizability of these findings to other treatment regimens may be limited. The retrospective nature of the study might introduce biases. Furthermore, the study mainly focused on overall survival and progression-free survival; other clinical endpoints could be explored in future studies. Finally, external validation in diverse populations would further strengthen the clinical applicability of this model.