Abstract
Introduction: Genetic classification of diffuse large B-cell lymphoma (DLBCL) using LymphGen (LG) has provided critical insights into DLBCL biology and informed precision medicine strategies aimed at improving patient outcomes. However, clinical application of LG remains limited due to cost, complexity, and an inability to classify a large proportion of DLBCLs (40-75%). Moreover, LG fails to incorporate non-genetic features impacting disease biology. To address these limitations, we developed LymphGen-sig (LG-sig), a gene expression platform that extends conventional LG classification. To evaluate its clinical utility, we assessed the impact of LG-Sig classification on progression-free survival (PFS) in the POLARIX study (NCT03274492).
Methods: LG-sig was developed using DLBCLs with paired genomic and transcriptomic data (NCI, n=481; BCC, n=283). Model development was restricted to DLBCLs classified by LG (v2.0; ≥90% probability) into MCD, BN2, EZB, or ST2. The dataset (n=276) was partitioned into training (75%) and validation (25%) for model development and performance assessment. Gene features were selected by differential gene expression, and a feature set of the highest ranked 294 genes was identified as optimal for classification using a nearest shrunken centroid classifier tuned with 5-fold cross-validation. LG-sig classifications were designated as MCDsig, BN2sig, ST2sig, and EZBsig. Within EZBsig, the dark zone signature was applied to classify cases as EZB-MYC+sig and EZB-MYC-sig. The final model was applied to RNAseq from archival tumors in POLARIX (n=678) to assess the PFS benefit of polatuzumab vedotin-R-CHP (pola-R-CHP) over R-CHOP for each LG-sig subtype. No adjustments were made for multiple comparisons.
Results: LG-sig classified conventional LG subtypes in both training and validation datasets with 84% and 80% balanced accuracy, respectively, which indicated that transcriptional features are sufficient to classify DLBCLs into conventional LG categories. Given the strong concordance of LG and LG-sig classifications, we examined whether LG-unclassified (LG-u) DLBCLs shared transcriptional features with specific LG-sig subtypes. LG-u cases were reclassified by LG-sig into MCDsig (28%); BN2sig (36%); ST2sig (22%); EZB-MYC+sig (5%); EZB-MYC-sig (9%). Reclassified LG-u cases exhibited broad transcriptional similarities to their genetically defined counterpart (e.g. MCDsig vs conventional MCD). Thus, despite often lacking canonical genetic drivers, LG-u DLBCLs retain LG subtype-specific transcriptional programs, underscoring the value of transcriptional profiling in extending genetic DLBCL classification.
In addition, when compared to other LG subtypes, striking intra-cluster heterogeneity was observed within LG A53 DLBCLs, which precluded establishment of a reproducible A53 signature via LG-sig. Instead, A53 DLBCLs were reclassified into other LG-sig subtypes (MCDsig 60%; BN2sig 18%; ST2sig 13%; EZB-MYC+sig 1%; EZB-MYC-sig 8%). As with LG-u cases, reclassified A53 DLBCLs had strong transcriptional similarity to other members of the same cluster, suggesting that A53 LG DLBCLs share greater biological similarity with other LG-sig subtypes than with each other – a finding that supports restructuring the A53 LG subtype.
When applied to POLARIX, LG-sig maintained high balanced accuracy (78%), which varied by subtype (EZB-MYC+sig: 92%, MCDsig: 83%, BN2sig: 81%, ST2sig: 69%, EZB-MYC-sig: 64%). Although no conventional LG subset previously showed a significant PFS benefit with pola-R-CHP over R-CHOP, MCDsig DLBCLs had significantly improved PFS with pola-R-CHP (HR 0.41, 95% CI: 0.26-0.66), even after adjusting for IPI, age, and cell-of-origin (HR 0.36, 95% CI: 0.22-0.59). Interestingly, there were trends towards benefit with R-CHOP over pola-R-CHP for ST2sig. Among BN2sig, there was potential benefit with pola-R-CHP, but only among ABC cases. Notably, within LG-u cases, only DLBCLs classified as MCDsig derived a PFS benefit with pola-R-CHP (HR 0.39; 95% CI: 0.23-0.67), highlighting the clinical relevance of transcriptional classification.
Conclusion: LymphGen-sig is a novel gene expression-based platform that complements conventional genetic classification by 1) expanding subtype assignment to include previously unclassified cases, 2) reassigning A53 DLBCLs into potentially more biologically relevant subtypes, and 3) uncovering a subset of LymphGen-unclassified DLBCLs that derive benefit from pola-R-CHP treatment.
