Chemical language models can generate biomolecules atom-by-atom

Abstract

Chemical language models are powerful generative models, capable of learning complex molecular distributions such as the largest molecules in Pubchem. In this work, we further show that chemical language models can learn atom-level representations of substantially larger molecules – scaling even to biomolecules like proteins. We show that chemical language models can generate entire biomolecules atom by atom – effectively learning the multiple hierarchical layers of molecular information from primary sequence to tertiary structure. Even further, we demonstrate that chemical language models can explore chemical space and protein space simultaneously by generating novel examples of protein-drug conjugates. The results demonstrate the potential for atom level biomolecular design with chemical language models.