AI Model CoCoGraph Generates Millions of Viable Molecules, Fooling Experts 40% of the Time

Context The number of possible molecules is estimated at up to 10^60, far exceeding the count of water molecules in Earth’s oceans. Searching this immense space for useful compounds is a major challenge in chemistry, from pharmaceuticals to sustainable materials. Researchers at Universitat Rovira i Virgili built CoCoGraph to explore this space by producing molecules that obey chemical laws.

Key Facts CoCoGraph operates as a diffusion model, a technique common in image generation, learning to reconstruct molecules after deliberately breaking their bonds. In a validation test, 121 chemistry experts viewed pairs of real and AI‑generated molecules and incorrectly identified the synthetic ones as genuine in roughly 4 out of 10 cases. The system guarantees 100% chemical validity by enforcing correct bond counts for each atom, unlike other models that can produce impossible structures.

What It Means Because CoCoGraph can quickly produce large libraries of realistic molecules, it may shorten early‑stage screening for new drugs or functional materials. Its ability to mimic real chemistry could also help scientists imagine compounds that might exist under extraterrestrial conditions. The next step is teaching the model to design molecules with targeted properties, such as specific solubility or biological activity.

Watch for efforts to link CoCoGraph’s output to experimental synthesis and functional testing.