Refining diffuse large B-cell lymphoma subgroups using integrated analysis of molecular profiles

Abstract

Background: Gene expression profiling (GEP), next-generation sequencing (NGS) and copy number variation (CNV) analysis have led to an increasingly detailed characterization of the genomic profiles of DLBCL. The aim of this study was to perform a fully integrated analysis of mutational, genomic, and expression profiles to refine DLBCL subtypes. A comparison of our model with two recently published integrative DLBCL classifiers was carried out, in order to best reflect the current state of genomic subtypes.

Methods: 223 patients with de novo DLBCL from the prospective, multicenter and randomized LNH-03B LYSA clinical trials were included. GEP data was obtained using Affymetrix GeneChip arrays, mutational profiles were established by Lymphopanel NGS targeting 34 key genes, CNV analysis was obtained by array CGH, and FISH and IHC were performed. Unsupervised independent component analysis (ICA) was applied to GEP data and integrated analysis of multi-level molecular data associated with each component (gene signature) was performed.

Findings: ICA identified 38 components reflecting transcriptomic variability across our DLBCL cohort. Many of the components were closely related to well-known DLBCL features such as cell-of-origin, stromal and MYC signatures. A component linked to gain of 19q13 locus, among other genomic alterations, was significantly correlated with poor OS and PFS. Through this integrated analysis, a high degree of heterogeneity was highlighted among previously described DLBCL subtypes.

Interpretation: The results of this integrated analysis enable a global and multi-level view of DLBCL, as well as improve our understanding of DLBCL subgroups.

Keywords: Diffuse large B-cell lymphoma; Gene signatures, prognosis; Independent component analysis; Transcriptomic variability.

Fig. 1

Component expression associations with mutations, FISH and IHC. Density plots are used in each cell. The X axis represents component expression level and the Y axis represents patient distribution (gaussian kernel density estimate). Red and blue respectively indicate the distribution of patients negative or positive for mutation, FISH or IHC as noted in cell title. Grey areas correspond to the overlap between the two distributions. Only plots where ICA expression is highly significantly different between positive and negative patients are shown here (Mann-Whitney, FDR<0.01). Significance of the difference is furthermore described by stars in the top right corner of each cell : ** for FDR < 0.01, *** for FDR < 0.001, **** for FDR < 0.0001, ***** for FDR < 0.00001. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 2

Graphic representation of component and chromosomal alteration associations. Chromosomes 1 to 22 are represented along with their cytogenetic bands. Proportions of the series presenting copy gain (blue) or loss (red) are presented as histograms around the chromosome ideogram. Significant associations of chromosomal regions with component expression levels are shown underneath each chromosome: yellow indicates positive associations and black indicates negative associations, with correlation intensity being depicted by variation in color intensity. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 3

Components with significant prognostic impact. Fig. 3a.Positive prognostic impact of component 11. Fig. 3b.Poor prognostic impact of component 23. Survival analyses were performed according to treatment regimen: R-chemo includes all patients with Rituximab regardless of associated chemotherapy, R-CHOP includes patients with R-CHOP and R-CHOP like regimens (R-miniCHOP), R-ACVBP includes patients with R-ACVBP treatment. Patients were grouped into tertiles relative to component expression levels (low, intermediate and high). Progression-free survival analyses are shown in the upper panels; overall survival analyses are shown in the lower panels. FDR of continuous Cox model applied to R-CHEMO is shown. P-value is calculated for high tertile vs low tertile with hazard ratio (HR) indicated.

Fig. 4

Application of genClass Algorithm to Cohort. Patients are distributed in columns and are grouped according to the genClass classification, cell of origin (COO) and NMF classification. Black squares highlight MCD, BN2, N1, EZB and STS clusters. Single Nucleotide Variants (SNV) and Structural Variants (SV) selected by the genClass algorithm are displayed in rows (one row per gene). The total number of each alteration is shown in the right-hand panel. “.” indicates event types which are disregarded by the algorithm. Gray cell background highlights the availability of CGH or FISH data for the considered patient, only for genes (rows) where it is relevant. For SV, “Wild-type” and “NA” columns correspond respectively to patients wild-type for the chromosomal aberration in question or without SV data.

Fig. 5

Application of NMF algorithm to cohort. Patients are distributed in columns and are grouped according to the NMF classification, cell of origin (COO) and genClass classification. Black squares highlight clusters A (CA), B (CB), C (CC) and D (CD). Single Nucleotide Variants (SNV) and Structural Variants (SV) selected by the NMF algorithm are displayed in rows (one row per event). The total number of each alteration is shown in the right-hand panel. “.” indicates event types which are disregarded by the algorithm. Gray cell background highlights the availability of CGH or FISH data for the considered patient, only for genes (rows) where it is relevant. For SV, “Wild-type” and “NA” columns correspond respectively to patients wild-type for the chromosomal aberration in question or without SV data.

Fig. 6

Relationship between genClass, NMF, COO and ICA classifications. Heatmap illustrating the interplay between genClass clusters (BN2, EZB, MCD, N1, STS and other), NMF clusters (CA, CB, CC, CD, CU), COO (ABC, GC, PMBL, other) and ICA-defined component expression level. Percentages within each cell of the heatmap indicate the proportion of patients within the cluster indicated in row that also belong to the cluster described in column. For example, 90% of the 30 EZB patients are clustered within NMF cluster CA while the remaining 10% belong to CC. Percentages on top of each classification-related column indicate the proportion of patients within each group considering the whole-cohort. For example, COO defines 38% of the cohort as ABC, 38% as GC, 8% as PMBL and 15% as “other”. Cell background highlights imbalances between these proportions, which would be expected in the absence of correlation, and the observed proportions (enrichments in red, depletions in blue). Color intensity varies according to the significance of a Fisher test (refer to the graphical legend at the bottom of the figure). Cell colors in the ICA panel vary in a similar fashion: a red background indicates a higher ICA score in the considered group as compared to the rest of the patients, while a blue background indicates a lower ICA score. Color intensity varies according to the significance of a Mann-Whitney test, with the same legend described at the bottom of the figure. ICAs are ordered from the highest number of significant correlations (on the left) to the fewest (on the right), with their numeric ID printed at the top of the column. Corresponding p-values and FDR are presented in Suppl Table 8. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)