High throughput screening (HTS) is a common starting point for small molecule drug discovery campaigns, especially when the biological target is new. HTS is commonly conducted in the wet lab (e.g. the U-M Center for Chemical Genomics), but also can be conducted in silico. These "virtual screens" are either structure-based, requiring a knowledge of the 3D structure of the biological target, or ligand-based, in which case having multiple overlapping ligands makes it easier to define the feature set that new ligands must match. In this talk, I will describe several applications of artificial intelligence/machine learning to virtual screening in our projects seeking to identify new starting points for cancer drug discovery. In so doing, I will trace the evolution of these methods over the last decade from crude attempts at generative design to the application of frontier models like Claude Science.