Identifying molecular features associated with anticancer drug response through pharmacogenomic analysis is the cornerstone of biomarker discovery. While learning from patients and tumor biopsies is the best approach to develop companion diagnostics that are the most likely to be clinically relevant, generating such data in clinical trials is lengthy and resource-intensive. In this context, preclinical models are important tools to study the complex relationship between molecular aberrations in cancer cells and response to genetic and chemical perturbations, as they allowed to generate large-scale “phenogenomics” datasets amenable to advanced computational analysis. However, it is well established that preclinical models do not fully recapitulate the drug response observed in patients due to their intrinsic limitations. Moreover, the inevitable noise in the genotype and phenotypic measurements due to the complexity of high-throughput experimental protocols makes biomarker discovery the more challenging. In this presentation, I will review the work from our group and others quantifying limitations of in vitro preclinical cancer models and the noise in phenogenomic data, and possible avenues to overcome this limitations in the computational analysis of these valuable data.