Parsimonious Gene Correlation Network Analysis in DLBCL Allows Refined Linkage to Mutation State and Treatment Response

Background

DLBCL can be divided into categories based on gene expression signatures. The two common classifications, Cell of Origin (COO) and Consensus Clustering Classification (CCC), assess overlapping but distinct aspects of DLBCL biology. We sought to explore the DLBCL expression landscape, without using prior knowledge, using a novel parsimonious gene correlation network approach (PGCNA).

Methods

Using 10 publicly available DLBCL data-sets we calculated pairwise correlations amongst the most variant genes. To aid module discovery the median correlations (across data-sets) were strongly filtered to remove noise. Resulting correlations were used to carry out 10,000 independent clusterings of the data utilizing a fast modularity optimisation method. Clusterings were ranked by modularity score and the top 100 clusterings were analysed for linkage of each module per cluster to biological function with a weighting to reward purity. Modules in each clustering were compared to calculate cluster stability. This process resulted in an optimised network level view of gene expression correlations in DLBCL.

Using sets of the most correlated (hub) genes per module, the enrichment of the modules was mapped back onto expression data-sets allowing classification of individual cases. The network modules were then used to assess the independent REMoDL-B DLBCL data-set (n=923, frontline study R-CHOP versus R-CHOP with bortezomib)

Results

The parsimonious gene correlation network approach (PGCNA) defines 26 modules corresponding to core biological processes and differentiation states including ABC-like, GCB-like, immune, lymph node, cell-cycle and metabolic. This integrated view, derived solely from the intrinsic structure of gene expression across many clusterings and recurrent across all data-sets, extends fundamental concepts of DLBCL disease biology and visualises the fundamental linkages between core biological processes. Analysis of the modules across the 11 data-sets reveals that the ABC- and GCB-like modules explain most of the variance across the DLBCL patients alongside the host-response/immunogenic cluster.

The association of modules with mutation status, assigned COO class (ABC, GCB and COO-UNC) and outcome was then tested in the REMoDL-B data-set. This established the significant enrichment of the ABC-like module with MYD88^***, CD79B^*** and PRDM1^*** and GCB-like with EZH2^***, CREBBP^***, MEF2B^*** and KMT2D^** mutations, confirming the association of these expression states with specific mutations independent of conventional COO classification. Novel associations in DLBCL between the cell-cycle module and DDX3X^***, the lymph-monocyte module and FAS^** and B2M^*, and the lymph-node/stromal module and TNFAIP3^** were also identified (p -value *** < 0.001,** < 0.01, *< 0.05). Interestingly more diverse associations are observed between selected gene mutations and expression modules, for example for MYD88, underlining the importance of defining both expression and mutation state.

Analysis of REMoDL-B survival data (OS and PFS) confirmed the association of differentiation-linked expression states with outcome trends independently of conventional COO-classification: ABC-like module with increased hazard ratio (HR), GCB-like module with decreased HR. Lymph-node and mast-cell modules were associated with decreased HR in both OS/PFS in both treatment arms (R-CHOP/RB-CHOP). Of note some modules differed in HR association between treatment types, for example the glycolysis module, associated with a decreased HR only in the R-CHOP treatment arm.

Conclusion

Network analysis provides an integrated view of the DLBCL expression landscape without a priori concepts. Discovery of expression modules corresponding to ABC and GCB-like states alongside those corresponding to OxPhos, BCR and Host Response groups provides a unified view of these features of DLBCL biology. The DLBCL expression modules allow refined linkage to mutation state and treatment response.

Figure

View largeDownload slide

Figure

View largeDownload slide

Close modal