A targeted gene signature stratifying mediastinal gray zone lymphoma into classical Hodgkin lymphoma-like or primary mediastinal B-cell lymphoma-like subtypes

Mediastinal gray zone lymphoma (MGZL), a B-cell lymphoma with overlapping features between primary mediastinal B-cell lymphoma (PMBL) and classical Hodgkin lymphoma (CHL), is a unique entity and a diagnostic challenge.^1,2 MGZL typically exhibits discordant morphological and immunophenotypic traits and a molecular profile straddling between PMBL and CHL.^3,4 However, the diagnosis of MGZL largely relies on morphological/immunophenotypic criteria⁵ and unstandardized connotations such as CHL- or PMBL-like entities may affect therapeutic choices and patients’ outcome. Retrospective studies revealed common reclassification of diagnosis and heterogeneous treatments associated with high relapse rates even following intensified chemotherapy,^6,7 emphasizing the need for new tools to improve the pathobiological stratification of MGZL. Here, we report the development of a signature - comprising both tumor- and tumor microenvironment (TME)-related genes - that enables MGZL categorization based on transcriptomic proximity to either CHL or PMBL. The study, conducted in line with the Declaration of Helsinki and with formal ethical approval (Comitato Etico Regionale per la Sperimentazione Clinica della Toscana with protocol number bioGZL-2020, Rif. CEAVC Em. 2022-263, Study number 18236_oss, 21/06/2022), was designed as reported in Figure 1.

With the speculative idea of a molecular allocation of MGZL between CHL and PMBL,^3,4 we first sought to identify a unique set of transcripts capable of discerning these two entities, considering potential unbalanced contributions to gene expression offered by tumor and TME cells. A discovery cohort comprising 84 CHL and 51 PMBL, further subdivided into training (50 CHL and 31 PMBL) and testing (34 CHL and 20 PMBL) sets, was generated by pooling Affymetrix-HG-133plus2 raw data from three different gene expression profile (GEP) datasets of fresh-frozen biopsy tissues (GSE17920, GSE11318, GSE87371),^8-10 and processed as two distinct expression matrices (data not shown). A combination of CIBERSORTx deconvolution (http://cibersortx.stanford.edu)¹¹ and a non-negative matrix factorization (NMF)-based approach¹² was used to identify gene sets that more accurately discriminate CHL from PMBL in an unsupervised fashion. CIBERSORTx was applied to create a customized signature matrix, including GEP of both tumor (N=2) and TME cytotypes (N=22) (Figure 2), used to derive two purified GEP matrices from the bulk transcriptome of the training cohort. Each purified matrix was independently decomposed by the NMF algorithm.¹² The method was run 100 times varying the rank value in the interval [2,7].

We adopted a cophenetic correlation coefficient (CCC) and consensus matrices to choose the optimal value of rank r for the factorization process. We choose as optimal rank the first value of rank r at which the CCC trend starts decreasing, and the one associated with clear block diagonal patterns in the consensus plots (r=2) (Online Supplementary Figure S1A, B, upper panels). NMF enabled an unsupervised selection of genes related to tumor (N=47 for CHL and N=653 for PMBL) and the TME (N=1,594 for CHL and N=637 for PMBL) (Online Supplementary Figure S1C, D). We consistently observed a considerable abundance of TME genes in CHL, whereas tumor-derived transcripts prevailed in PMBL. Unsupervised clustering highlighted the capacity of these genes, merged in a unique panel of 2,913 transcripts (data not shown), to fully and reproducibly separate CHL from PMCL (Online Supplementary Figure S1C, D). By further filter-based feature selection (ReliefF¹³ and Laplacian score¹⁴), we selected a final panel of 168 genes, which retained this discriminative ability both in the training and testing cohorts (Online Supplementary Figure 2A, B). Interestingly, genes associated with T-cell receptor signaling (e.g., CD28 and CD3G), inflammation (e.g., PRDX2) and STAT5A targets (e.g., PRTFDC1) stood out as related to CHL, whereas genes involved in the cell cycle (e.g., BTRC, MCM6, SPC25 and RNF8), chromatin-modifying enzymes (e.g., HCFC1), and those involved in the regulation of TP53 activity (e.g., TP63) emerged as associated with PMBL, suggesting a peculiar enrichment of known pathways for each disease. A 168-code set (including 15 housekeeping genes) was then customized to digitally profile (NanoString nCounter Flex Analysis System, NanoString Technologies) a real-life independent collection of CHL, PMBL, and MGZL cases whose RNA was directly extracted from formalin-fixed paraffin-embedded specimens (MagMAX FFPE RNA/DNA Ultra Kit, ThermoFisher Scientific). Under the auspices of the Italian Lymphoma Foundation, a multicenter collection of 39 cases originally diagnosed as gray zone lymphoma underwent central pathology review by a panel of expert hematopathologists (ES, CA, SAP, ML, MP, LL, FF, AD, ML, MP, LL, SL, SA, AS and AZ) to assess whether these fulfilled the current World Health Organization classification¹ and International Consensus Classification² by hematoxylin/ eosin and immunohistochemical staining, and were also classified according to Sarkozy et al.⁵ In situ hybridization for Epstein-Barr virus was performed and only negative cases with mediastinal involvement were considered.

Among the 39 cases with an original diagnosis of gray zone lymphoma, 28 were confirmed as MGZL and only 24 passed the quality check for the final study phase. Their histopathological features were defined as closer to CHL (CHL-like), PMBL (PMBL-like) or intermediate (Online Supplementary Figure S2D), also according to the morpho-phenotypic subgroups of Sarkozy et al.⁵ (Online Supplementary Table S1). Eight cases (Sarkozy group 0 [N=2], group 1 [N=5] and group 2 [N=1]) showed CHL-like morphology with Hodgkin and Reed-Sternberg cells within an inflammatory background and a variable degree of fibrosis, associated with sheets of monomorphic mononuclear medium/large cells. Ten cases (Sarkozy group 1/group 2 [N=2], group 2 [N=6] and group 2/group 3 [N=2]) showed PMBL-like morphology with a predominance of medium/large tumor cells mixed with a variable number of Reed-Sternberg-like cells, and low inflammatory/fibrosis background. These cases were CD30-positive with variable expression of CD15. Two cases were negative for CD20 and CD79A with weak to moderate expression of PAX5. Other cases displayed incomplete expression of B-cell markers, a moderate inflammatory background, and variable expression CD30. The remaining six cases (Sarkozy group 1/group 2 [N=5] and group 2/ group 3 [N=1]) displayed intermediate features, with sheets of mononuclear cells expressing of CD30 extensively, and a predominant full B-cell phenotype. These cases, on a pathological ground, were very challenging to categorize due their intermediate morpho-phenotype, even according to the Sarkozy classification.

Figure 1.Schematic overview of the study design. The methodological workflow included three different phases and independent cohorts of patients. The gene selection phase related to the training/ testing set was conducted in-silico (light blue), while the validation (light green) and MGZL assignment phases (red) were completed on a real-life set of cases (24 out of 28 centrally revised MGZL passed the quality control for the final study phase). CHL: classical Hodgkin lymphoma; PMBL: primary mediastinal B-cell lymphoma; GZL: gray zone lymphoma; MGZL: mediastinal gray zone lymphoma.

By applying our 168-gene signature on the NanoString platform to a real-life set of CHL (N=18) and PMBL (N=19), we first confirmed its capability to fully distinguish the two diseases (Online Supplementary Figure 2C). Then, the signature was used to stratify 24 MGZL based on their transcriptional proximity to either CHL or PMBL. As illustrated in Online Supplementary Figure S3, seven out of eight cases (pathologically annotated as CHL-like and in accordance with the Sarkozy subgroup) consistently clustered within the CHL subgroup, whereas only one showed a discordant allocation. Conversely, out of the ten PMBL-like MGZL, three fell concordantly within PMBL, whereas seven cases located within the CHL cluster. The three MGZL that consistently clustered into the PMBL subgroup shared strong PAX5 and moderate CD30 staining, with a more complete B-cell phenotype. In the remaining seven MGZL, PAX5 staining was weak or moderate, CD30 at strong intensity and other B-cell markers variably positive. The remaining six cases - with intermediate morphology/Sarkozy assignment – mostly clustered within CHL, while one was within the PMBL cluster. The transcriptional clustering resulted in a valuable separation of cases, mostly falling into the CHL subgroup, especially those with intermediate morphology and Sarkozy group 1/group 2, although a subset of cases unanimously classified as PMBL-like also clustered in this group. In line with previous reports,⁴ 19 out of 24 MGZL in our study appeared transcriptionally closer to CHL (Online Supplementary Table S1). The availability of clinical information for a subset of 14 MGZL prompted us to speculate on the hypothetical relationship between therapeutic outcome and pathological/molecular stratification of cases (Online Supplementary Table S2). A complete response was recorded for eight cases - all treated with R-CHOP (rituximab plus cyclophosphamide, doxorubicin, vincristine, prednisone)-like regimens - of which three were pathologically and molecularly assigned to the CHL cluster, three were PMBL-like including only one case in molecular concordance, and two cases had intermediate morphology, one of which fell within the CHL cluster and the other within the PMBL cluster. Intriguingly, six MGZL showing poorer outcomes were mostly characterized by a discrepancy between transcriptomic clustering and treatment. In particular, one case - pathologically annotated as CHL-like and concordantly assigned to the CHL cluster - reached a partial response after a typical PMBL-oriented protocol (da-EPOCH-R: dose adjusted etoposide, prednisone, vincristine, cyclophosphamide, doxorubicin, and rituximab). Likewise, one case within the PMBL cluster but characterized pathologically as CHL-like was treated with a CHL-oriented therapy (ABVD: adriamycin, bleomycin, vinblastine, dacarbazine) and displayed refractoriness. The other three PMBL-like MGZL - but molecularly allocated into the CHL cluster - exhibited unfavorable outcomes after induction with R-CHOP-like protocols. A partial response was also recorded for a PMBL-like case (both morphologically and molecularly) treated with da-EPOCH-R. Although obtained by an unprecedented approach, our results strengthen the idea that the majority of MGZL are biologically closer to CHL,⁴ reinforcing the value of previous observations^15,16 on the efficacy of anti-CD30/anti-PD1 drugs in MGZL, although as salvage strategy after PMBL-oriented frontline treatments. More accurate MGZL categorization, which to date remains heterogeneous and guided solely by the morpho-phenotypic depiction of the disease, may prompt the development of new frontline combinations of drugs.

Figure 2.Identification of a molecular signature distinguishing classical Hodgkin lymphoma and primary mediastinal B-cell lymphoma. Schematic overview of the CIBERSORTx analysis. We applied CIBERSORTx for digital purification of a bulk mixture matrix (50 classical Hodgkin lymphomas and 31 primary mediastinal B-cell lymphomas as the training set) using a customized signature matrix of 24-cell types, composed of two tumor cytotypes and 22 tumor microenvironment (TME) cells. Tumor and TME compartments were run as merged classes to purify the gene expression profile for both components. The t-distributed stochastic neighbor embedding visualization of the two purified tumor and TME gene expression profiles are depicted on the right side. CHL: classical Hodgkin lymphoma; PMBL: primary mediastinal B-cell lymphoma; BC: B cells; TC: T cells; NK: natural killer cells; Mo: macrophages; DC: dendritic cells; GEP: gene expression profiles; t-SNE: t-distributed stochastic neighbor embedding.

In conclusion, aiming at laying the basis for a more accurate stratification of MGZL, we diversified our approach from previous studies^3,4 by exploiting computational tools to obtain a restricted gene panel - easily measurable on RNA from formalin-fixed paraffin-embedded samples - which fully separates CHL from PMBL, and molecularly assigns MGZL to one or other of the entities. Such an approach might help in overcoming the histopathological challenge of MGZL categorization, despite efforts to apply the Sarkozy classification to describe proximity to CHL or PMBL. Validation of our signature on larger, independent sets of MGZL would be critical to decipher the underlying relationship between molecular and phenotypic traits, to build a combined histopathological/transcriptomic model of MGZL stratification and, ultimately, to prompt its translation into the clinical setting in order to optimize the treatment of these rare cases.

Footnotes

• Received February 22, 2024
• Accepted July 8, 2024

Correspondence

*GG and MCV contributed equally as first authors

S. CIAVARELLA - s.ciavarella@oncologico.bari.it

Disclosures

No conflicts of interest to disclose.

Funding

Acknowledgments

GG, FE, LS and NDB are members of the Gruppo Nazionale Calcolo Scientifico - Istituto Nazionale di Alta Matematica (GNCS-INdAM).