Blank Infilling for Generalist Chemical Language Models

Drug design requires exploring a vast chemical space under multiple property constraints, yet most generative models remain specialized to individual tasks. We introduce OmniInvent, a generalist chemical language model that operates directly on unmodified SMILES strings and supports both de novo and substructure-constrained molecular generation. OmniInvent formulates molecular design as autoregressive blank infilling over partially masked SMILES, using prefix attention, two-dimensional positional embeddings, and a graph-based masking strategy for connected molecular substructures. In goal-directed design, OmniInvent achieved best-in-class performance on the Practical Molecular Optimization benchmark for de novo design and matched specialized state-of-the-art models in substructure-constrained settings. These results position blank infilling on SMILES as a strong foundation for drug design.