Creation of a Gene Expression Signature-Based Prediction Model Specific to Proteasome Inhibitor Response in Multiple Myeloma

Multiple myeloma (MM) is the second-most common hematopoietic malignancy in the United States with significant complexity and heterogeneity at the molecular level. Proteasome inhibitors (PIs) including Bortezomib/Bz (Velcade), carfilzomib/Cz (Krypolis) and recent orally active second generation PIs MLN9708 (Ixazomib/Ix) and Oprozomib/Opz are becoming effective chemotherapeutic agents in the treatment of MM, used alone or in combination with other anti-cancer agents. However, despite these recent improvements in treatment strategies, MM still remains an incurable disease with median survival rate of around 7 years. Wide inter-individual variation in response/resistance to PI treatment is a major limitation in achieving consistent therapeutic effect in MM. Such heterogeneity in response to PI-based treatment is likely governed by the underlying molecular characteristics of the MM tumor cells including genomic and transcriptomic alterations.

In the current study, we used in vitro chemosensitivity data to develop a gene expression-based model to predict therapeutic response to proteasome inhibitors that were cross-validated in other cell-based MM models and clinical trials on MM patients. A panel of 50 human multiple myeloma cell lines (HMCLs) representing the biological and genetic heterogeneity of MM was used to generate single-agent cytotoxicity profiles for 4 the PIs Bz, Cz, Ix and Opz based on half maximal inhibitory concentration (IC₅₀) values. Gene expression profiling (GEP) was performed using llumina's HiSeq 2000 next-generation high-throughput sequencing technology (RNAseq) for 50 paired-end reads with depth of >20 million reads per sample.

Subsequently, GEP data was used to perform relevant gene selection and eventually develop a gene-based prognostic prediction model associated with PI response. In terms of statistical methodology, we used a locally additive polynomial regression model, which is a nonparametric regression model, on multivariate responses generated by our data. This was coupled with a sparsity-driven statistical model for dimension reduction and grouping of GEP data corresponding to drug response of the different cell-lines.

Our GEP-based prediction model could successfully distinguish between good or poor PI response with high level of significance (p<0.005) and stratify survival in the PI-treatment arms of two independent sets of clinical trials on MM patients, APEX (Hazards ratio (HR) = 2.185; p=0.001) and UAMS-MMTT3a (HR= 2.091; p=0.009) but not for the non-PI arms suggesting PI-specificity of our prediction model.

Therefore, we could successfully generate a gene expression signature-based prediction model qualitatively and quantitatively associated with PI response that could be effectively cross-validated on in vitro tumor models and human MM clinical trials. Our research will benefit clinical decision-making through the pre-identification of non-responders to PI treatment prior to initiating MM therapy and the development of novel treatment strategies specifically targeted at PI-resistant MM patients.