Hot off the press – excited to reshare this new preprint from the Polaris - Benchmarks for methods that matter consortium addressing replicability challenges in ML-based #drugdiscovery!? ? Experts from leading pharma and biotech companies propose clear guidelines for evaluating ML methods, an essential step toward bridging the gap between perceived progress and real-world impact. We’re proud that our VP of Computational Chemistry and Structural Biology, Daniel Price, is a co-author of this important work!? https://lnkd.in/g6w78h9W ? #DrugDiscovery #MachineLearning?
Introducing our first proposed set of guidelines for method comparison in small molecule property prediction! ML research faces replicability and reliability issues around benchmarking of new methods. This is particularly true of small molecule predictive modelling tasks, where costly decisions rely on noisy, imbalanced?data. Crafted by the Small Molecule Steering Committee, these guidelines introduce statistically rigorous, domain-appropriate comparison protocols for small molecule predictive modelling to help ensure replicability and practical impact. The field needs better ways to compare models to help us understand how we’re progressing and which models to deploy in real-world drug discovery scenarios. The pre-print provides recommendations on performance sampling distribution, statistical testing, emphasizing practical significance, and showcasing your results beyond a traditional leaderboard. Read the pre-print for more details: https://lnkd.in/g6w78h9W Annotated examples of the protocol in action: https://lnkd.in/gMzpS4TZ Let’s work together to redefine how method comparison is done! We want to hear from you. Is there something we missed? Do you have other thoughts on the method comparison protocol? Connect with us on GitHub: https://lnkd.in/gMzpS4TZ