An Integrative Network Approach for Genomic Discovery and Personalized Therapy in Relapsed Multiple Myeloma

Novel methods are needed to leverage the complex information generated by recent large-scale genomic profiling efforts in Multiple Myeloma (MM) to inform therapeutic discovery and personalized medicine strategies. We propose that network modeling can be used to organize patient specific gene signatures in a way that is consistent with and builds on known molecular classifications of multiple myeloma, can facilitate identification of putatively novel sub-types, and can inform prioritization of therapies. In this study, we applied weighted gene co-expression analysis WGCNA (Langfelder, 2008) to 116 relapsed MM samples in the MMRC reference collection, producing 18 modules of co-expression genes. We identified significant correlations between several co-expression modules and clinical traits relevant to multiple myeloma (e.g. serum LDH, albumin).

To compare these modules with existing methods of disease classification, we used gene expression profiles to assign patients into one of the seven categories of molecular classification for multiple myeloma (Zhan, 2006). The classifier was built using a PAM approach (Hastie, 2006), achieving an accuracy of >95% in test datasets. Each of the samples used to build the co-expression network was classified, building up a map of associations between sample traits, co-expression modules, and molecular classifications. With the exception of Cyclin-D group-2 (CD-2), which was not associated with any modules, each of the molecular classification groups has a unique set of module associations.

We observed consistent agreement between functional molecular and clinical annotations of network modules. For example, we identified a network module associating with the Proliferative (PR) subtype (Zhan, 2006) of MM that was also positively enriched for genes located on the 1q locus of chromosome 1. Another module associated strongly with the c-MAF/MAFB subtype (MAF), was enriched for known targets of MAF, including ITGB7, CX3R1, NUAK1/ARK5, NTRK2, ARID5A, SMARCA1, TLR4, SPP1 and G6MB6.

To explore how the co-expression network could be used to identify therapeutic options, we projected each module on to a large library of drug induced transcriptional profiles (Lamb, 2006), identifying compounds predicted to perturb the module in a specific manner. To characterize the chemogenomic features that may underpin the relevance of these drugs to each module, we performed target set enrichment analysis for each module’s drug-list, building a map of the most relevant drug targets for each module. We found that drugs predicted to modulate the MAF associated module are known to target Adenosine receptor 2A (ADORA2A), which is not associated with any other module. Reports of ADORA2A agonists synergizing with dexamethasone and PDE inhibitors induce apoptosis and reduce proliferation in myeloma samples (Rickles, 2010), suggesting ADORA2A modulators might be relevant to the MAF subtype.

We assessed the ability of the network to organize individual patient transcriptional profiles using an independent set of RNA-seq profiles of CD138+ cells from twenty-eight multiple relapsed MM patients treated at Mount Sinai. Patient-specific differential gene expression was estimated by calculating the fraction of samples with a lower FPKM than a given sample. This vector of gene scores was used to perform a rank based measure of gene set enrichment against each of the modules using GAGE (Luo, 2009). Clustering the resulting sample-module scores using multiscale bootstrap resampling (Suzuki, 2006) uncovered 5 stable clusters. We applied the Zhan classifier to these samples and looked for enrichment within each stable clusters for each class. There were four novel clusters in addition to one cluster positively enriched (P < 0.003) for MS, a poor prognosis class, defined by MMSET-spiking, and overexpression of genes localizing to chr1q. Clustering these samples according to gene expression did not reproduce the MS (or any other) cluster enrichment associations, suggesting the potential of this network approach for organizing patient samples in a way that is clinically and biologically relevant.