The ML drug discovery startup trying really, really hard to not cheat (www.owlposting.com)

Leash Bio, a Utah-based startup specializing in machine learning for drug discovery, is making waves by emphasizing the importance of rigorous model validation in a field often marred by unintentional "cheating." Founded in 2021 by Ian Quigley and Andrew Blevins, the company aims to ensure its models genuinely learn useful information about small-molecule interactions instead of inadvertently subverting their efficacy by relying on biases in training data. This dedication is particularly significant as many existing models are criticized for simply memorizing previous data rather than generalizing effectively to new chemical entities. One of Leash Bio's innovative approaches involved issuing a public Kaggle challenge to address data biases—presenting a dataset of 133 million small molecules and inviting the community to rethink their methodologies. This initiative is a bold move to not only promote transparency but also to enhance the overall robustness of drug discovery models. By aligning with the broader AI/ML community's strive for integrity and reliability in machine-learning applications, Leash Bio's efforts may pave the way for advancements that could significantly improve the precision of drug design processes, thereby influencing the future of therapeutics development.