Generative AI models in chemistry are increasingly popular in the research community. They have applications in drug discovery and organic materials (solar cells, semi-conductors). Their goal is to generate virtual molecules with desired chemical properties (more details in this
blog post).
However, this flourishing literature still lacks a unified benchmark. Such benchmark would provide a common framework to evaluate and compare different generative models. Moreover, this would allow to formulate best practices for this emerging industry of ‘AI molecule generators’: how much training data is needed, for how long the model should be trained, and so on.
That’s what the
DiversityNet benchmark is about. DiversityNet continues the tradition of data science benchmarks, after the MoleculeNet benchmark (
Stanford) for predictive models in chemistry, and the ImageNet challenge (
Stanford) in computer vision.