Sampling a trillion-sized GDB-20 database of drug-like molecules by generative artificial intelligence

Identifying new chemical entities for small molecule drug discovery requires the continued exploration of its vast chemical space, which has been speculated to contain 10⁶⁰ or more molecules. Here we used generative artificial intelligence to sample a GDB-20 database estimated to contain 32 trillion organic molecules corresponding to 26.9 million molecular graphs up to 20 nodes selected for their limited topological complexity. We show that GDB-20 populates a distinct chemical space of sp³-rich, 3D-shaped and natural product-like molecules with favorable synthetic accessibility score. The relevance for drug discovery of our 12 billion sample, named GDB-20s and corresponding in size to approximately 0.04% of the estimated full database, is suggested by high compliance with drug-likeness and toxicity filters, and the coverage of thousands of scaffolds occurring in bioactive molecules recorded in the ChEMBL database. The generative models used to produce GDB-20s, together with the 12-billion-molecule collection are available for download at https://gdb.unibe.ch.