AI Tool Generates Millions of Molecules

Researchers at Spain's Universitat Rovira i Virgili unveiled CoCoGraph. This model, presented in research published in Nature Machine Intelligence, generates large numbers of new molecules not found in existing databases through a decompose-and-recombine approach using constrained discrete diffusion.

• It recorded 100% chemical validity, with uniqueness of 99.8%/99.9% and novelty of 95.7%.
• The BASE model has 534K parameters and the FPS model has 4.4M parameters, both smaller than many competing models, yet it achieved property matching KL divergence of 95.7%/96.3% on GuacaMol.
• Across 36 chemical properties, it also outperformed competing systems on at least 66.6% of the metrics.
• In a discrimination experiment involving 121 participants with backgrounds in chemistry and biochemistry, the accuracy of distinguishing real molecules from generated ones was 62%, close to random chance.

The researchers built a database of 8.2 million molecules and also demonstrated searching for paracetamol-like candidates and inpainting-style editing that fixes partial structures. However, the current stage does not yet directly design molecules for target functions, and constraints remain, including a maximum of 70 atoms and a fixed molecular formula.