Breakthroughs in machine learning theory and practice, coupled with ready access to cloud based supercomputing resources and ever-increasing amounts of exptl. data, are enabling truly AI centric processes for small mol. drug design wherein predictive models successfully substitute for laboratory assays throughout the Discovery critical path. This presentation will describe how diverse machine learning techniques, ranging from multidimensional and multitask boosting to deep neural networks, can extract accurate, scaffold independent, ligand based predictive models for important phenomena: target binding, functional activity, selectivity, PK/ADME properties, and toxicity. Applications of these models will be illustrated with examples from therapeutic programs and discussed in terms of their potential to enhance success/reduce attrition in drug discovery.