Just throw it away? Class imbalance lessons from molecular machine learning to meatballs

YouTube talk | GitHub repository One of the challenges in drug discovery is that most of the things you screen are not going to work. In antibiotic discovery, for example, you might test thousands of molecules in the lab and find that only a single digit percentage have the antibacterial activity you’re looking for (and that would be pretty impressive, honestly!). That’s a lot of negative data to collect. But do we really need all of it?