AbstractMany cancer drugs only benefit a subset of the patients that receive them, but are often associated with serious side effects. Predictive classification methods that can identify which patients will benefit from a specific treatment are therefore of great clinical utility. We here introduce a novel machine learning method to identify predictive gene expression signatures, based on the idea that patients who received different treatments but exhibit similar expression profiles can be used to model response to the alternative treatment. We use this method to predict proteasome inhibitor benefit in Multiple Myeloma (MM). In a dataset of 910 MM patients we identify a 14-gene expression signature that can successfully predict benefit to the proteasome inhibitor bortezomib, with a hazard ratio of 0.47 (p = 0.04) in class ‘benefit’, while in class ‘no benefit’ the hazard ratio is 0.91 (p = 0.68). Importantly, we observe a similar classification performance (HR class benefit = 0.46, p = 0.04) in an independent patient cohort which was moreover measured on a different platform, demonstrating the robustness of the signature. Moreover, we find that the genes in the discovered signature are essential, as no equivalent signature can be found when they are excluded from the analysis. Multiple genes in the signature are linked to working mechanisms of proteasome inhibitors or MM disease progression. In conclusion, our method allows for identification of gene expression signatures that can aid in treatment decisions for MM patients and provide insight into the biological mechanism behind treatment benefit.