Is GPT all you need for low-data discovery in chemistry?

Machine learning has revolutionized many fields and has recently found applications in chemistry and materials science. The small datasets commonly found in chemistry lead to various sophisticated machine-learning approaches that incorporate chemical knowledge for each application and therefore require a lot of expertise to develop. Here, we show that large language models that have been trained on vast amounts of text extracted from the internet can easily be adapted to solve various tasks in chemistry and materials science by simply prompting them with chemical questions in natural language. We compared this approach with dedicated machine-learning models for many applications spanning properties of molecules and materials to the yield of chemical reactions. Surprisingly, we find this approach performs comparable to or even outperforms the conventional techniques, particularly in the low data limit. In addition, by simply inverting the questions, we can even perform inverse design successfully. The high performance, especially for small data sets, combined with the ease of use, can have a fundamental impact on how we leverage machine learning in the chemical and material sciences. Next to a literature search, querying a foundational model might become a routine way to bootstrap a project by leveraging the collective knowledge encoded in these foundational models.