ExCAPE was a European funded project aiming at harvesting the power of supercomputers to speed up drug discovery (http://excape-h2020.eu/). Thanks to the project team, we were given the amazing opportunity to build large-scale machine learning models for compound activity predictions from public databases and to apply them to industrial datasets. In this talk, I will present the process of collecting chemogenomics data from public resources to build a benchmark dataset. Subsequently, I will explain the process of building and evaluating the performance of models built with multi-task deep learning and matrix factorization algorithms. Ultimately, I will show how these models were applied to industrial datasets.