Extensive interindividual variation in response to chemotherapy is a major stumbling block in achieving desirable efficacy in the treatment of cancers, including multiple myeloma (MM). In this study, our goal was to develop a gene expression signature that predicts response specific to proteasome inhibitor (PI) treatment in MM. Using a well-characterized panel of human myeloma cell lines (HMCLs) representing the biological and genetic heterogeneity of MM, we created an in vitro chemosensitivity profile in response to treatment with the four PIs bortezomib, carfilzomib, ixazomib and oprozomib as single agents. Gene expression profiling was performed using next-generation high-throughput RNA-sequencing. Applying machine learning-based computational approaches including the supervised ensemble learning methods Random forest and Random survival forest, we identified a 42-gene expression signature that could not only distinguish good and poor PI response in the HMCL panel, but could also be successfully applied to four different clinical data sets on MM patients undergoing PI-based chemotherapy to distinguish between extraordinary (good and poor) outcomes. Our results demonstrate the use of in vitro modeling and machine learning-based approaches to establish predictive biomarkers of response and resistance to drugs that may serve to better direct myeloma patient treatment options.