Repurposing FDA-approved drugs with the aid of gene signatures of disease can accelerate the development of new therapeutics. A major challenge to developing reliable drug predictions is heterogeneity. Different gene signatures of the same disease or drug treatment often show poor overlap across studies, as a consequence of both biological and technical variability, and this can affect the quality and reproducibility of computational drug predictions. Existing algorithms for signature-based drug repurposing use only individual signatures as input. But for many diseases, there are dozens of signatures in the public domain. Methods that exploit all available transcriptional knowledge on a disease should produce improved drug predictions. Here, we adapt an established meta-analysis framework to address the problem of drug repurposing using an ensemble of disease signatures. Our computational pipeline takes as input a collection of disease signatures, and outputs a list of drugs predicted to consistently reverse pathological gene changes. We apply our method to conduct the largest and most systematic repurposing study on lung cancer transcriptomes, using 21 signatures. We show that scaling up transcriptional knowledge significantly increases the reproducibility of top drug hits, from 44% to 78%. We extensively characterize drug hits in silico, demonstrating that they slow growth significantly in nine lung cancer cell lines from the NCI-60 collection, and identify CALM1 and PLA2G4A as promising drug targets for lung cancer. Our meta-analysis pipeline is general, and applicable to any disease context; it can be applied to improve the results of signature-based drug repurposing by leveraging the large number of disease signatures in the public domain.