Universitat Rovira i Virgili Researchers Create AI System to Generate Millions of Molecules

Researchers at Universitat Rovira i Virgili have developed an artificial intelligence model called CoCoGraph that can generate millions of new molecules that comply with chemical laws, according to the university. The study has been published in the journal Nature Machine Intelligence.

CoCoGraph works similarly to generative AI tools used for text or images. The system uses a diffusion model that progressively disorders a real molecule and then learns to reconstruct it. This allows it to produce chemically coherent structures without violating bonding rules. The researchers state that all molecules generated by the model are chemically valid, unlike some outputs from other models.

The research team compared CoCoGraph with other molecular generation systems by analyzing 36 physicochemical properties such as solubility and structural complexity. For about two thirds of these properties, the molecules generated by CoCoGraph were found to be more realistic. The model also uses fewer parameters and requires less computing power, improving efficiency.

To test plausibility, 121 chemistry experts were asked to distinguish between real molecules and those generated by the AI. They misidentified the artificial molecules 40 percent of the time, indicating that many generated compounds appeared authentic. The researchers also identified molecules with properties similar to paracetamol among the millions produced.

The team plans to extend the model to design molecules with specific functions, such as solubility or non-toxicity. They describe this as a first step toward AI systems capable of producing tailor-made compounds for use in chemistry, pharmacology, and materials science.