We developed a fully unsupervised deep learning method to cluster high-content cellular images thus extracting phenotypic similarities directly from fluorescence pixel intensity values. The method is based on two state-of-the-art methods, a neural network architecture developed at Novartis and a deep clustering framework developed at the Facebook AI Research. This combination allows us to investigate the cellular phenotypes arising from perturbations induced by various compound treatments. Comparison with anchor treatments, which phenotypic effects are known, subsequently allow us to investigate the mechanism of action of many treatments in a high-throughput fashion.