Hardwiring chemistry into diffusion to generate realistic molecules The space of drug-like molecules is roughly 10^60 candidates, and most generative models waste capacity learning what every chemistry student knows on day one: atoms have valences, bonds come in fixed multiplicities, not every graph is a molecule. The result is models that produce invalid structures or need millions of parameters to keep validity high. Manuel Ruiz-Botella and coauthors take a different route with CoCoGraph, a discrete graph diffusion model where chemical validity is built into the noising process rather than learned. Each step performs a double edge swap: bonds AB and CD are removed and replaced by AC and BD. This preserves the molecular formula and every atom's valence, so the trajectory never leaves the space of valid graphs. The denoiser only learns structural patterns, not chemistry's rules. A second idea is the collaborative mechanism. Some molecules randomize faster than others, so the nominal time step is a poor proxy for how far you are from a real molecule. A small auxiliary model estimates that distance from the graph and picks the trajectory point with the smallest predicted time. On GuacaMol (a standard benchmark scoring validity, novelty and property-distribution match against real molecules) against PubChem, CoCoGraph hits 100% validity, 95.7% novelty, and the best KL divergence, beating DiGress, GDSS, GruM and DeFoG with an order of magnitude fewer parameters. Across 36 descriptors picked by an external LLM to avoid cherry-picking, it wins on at least two thirds. The team also released 8.2M synthetic molecules and ran a Turing test with 121 chemists: postgraduates only reached 64% accuracy, and for acyclic or aliphatic molecules no better than random. This matters because validity-by-construction removes layers usually sitting downstream in drug and materials pipelines: post-hoc filtering and retraining cycles to push validity up. Combined with inpainting for scaffold preservation, it fits naturally into lead optimization, where teams grow fragments around a fixed core. Paper: Ruiz-Botella et al., Nature Machine Intelligence (2026) — journal license | doi.org/10.1038/s42256-026-0…