Abstract Patient-derived tumor cells are highly variable in their morphology and molecular phenotypes and after exposure to a library of drugs their heterogeneity further increases as the cells respond to the perturbations. The complete phenotypic diversity of the tumor cells and drug response patterns may remain undetectable if only few simple readouts, such as individual biomarkers in pre-defined tumor cells, are considered. Therefore, novel techniques that capture and visualize the entire spectrum of heterogeneous phenotypic changes in ex-vivo patient-derived tumor cells are needed. Here, we applied unsupervised machine learning for comprehensive visualization of cell phenotypes with the aim to improve the analysis of image-based drug screening results. We performed high content screening of patient-derived renal clear cell carcinoma cells, exposed to 36 drugs in nine concentrations applied in 384-well format. The nuclei were stained with Hoechst and cell proliferation using Ki67, and images were acquired with a high-content imaging microscope (PE Operetta, 20x). To quantitatively describe the cell phenotypes, we used CellProfiler for cell segmentation and feature extraction and then applied a large-scale data embedding method (LargeVis) to cluster the cells based on their phenotypes. Using the clustering, we created an image-based phenotype-map to describe the phenotypic landscape. Overlaying the drug treatments on different concentrations on top of the map allowed for visualization of unique phenotypic fingerprints for each drug at the single cell resolution. Visual inspection of the phenotype-map displayed visually similar cells clustering together. Likewise, a comparison with an expert trained supervised classification indicated consistency with the phenotype-map. As expected, Ki67-positive and -negative cells as well as dying cells formed clusters. Interestingly, the phenotype-map identified sub-clusters within the Ki67-positive and -negative cells. Furthermore, the map also allowed us to adjust for classification errors occurring in the supervised classification. In conclusion, we find that phenotypic signatures provide an intuitive way of linking and comparing phenotypes of distinct populations of drug-treated cells at a single-cell level. This will reveal additional systematic information of the drug responses, complementing traditional readout, such as cell viability, by highlighting the phenotypic changes over concentrations. Citation Format: Riku Turkki, Lassi Paavolainen, Piia Mikkonen, Päivi Östling, Peter Horvath, Vilja Pietiäinen, Olli Kallioniemi. Phenotypic heterogeneity of patient-derived tumor cells visualized by unsupervised analysis in cell-based personalized drug testing [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2018; 2018 Apr 14-18; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2018;78(13 Suppl):Abstract nr 5302.