CoCoGraph posts 100% validity for molecules

Molecule-generating AI has had a basic problem for years — it can be creative, but it also likes to invent nonsense. A model can spit out structures that look novel and score well on paper, then fail (nature.com)oGraph is interesting because it tries to remove that failure mode at the source, not patch it afterward. A team led by Manuel Ruiz-Botella, Roger Guim(nature.com)del generates molecules that are chemically valid by construction, and the paper just landed in Nature Machine Intelligence. (nature.com)sion model for molecules. Instead of treating a molecule like a text string, it treats atoms as nodes and bonds as edges, then learns how to reverse a controlled corruption process and rebuild plausible structures. The twist is that the model bakes chemical constraints into generation itself, so allowed moves are restricted by rules like proper bonding patterns rather than checked only at the end. (nature.com) ### Why is “100% validity” a big deal? Because validity is the floor, not the ceiling. If a generator cannot reliably produce chemically legal structu(nature.com)iction, and human evaluation all waste time on junk. The CoCoGraph paper claims 100% chemical validity on the benchmark it uses, which is a stronger claim than “better average quality” and closer to saying the model stops making a whole class of mistakes. (nature.com) ### How does it pull that off? Basically, the model collaborates with a second component that tracks how close the current graph is (nature.com)iary time model helps steer those edits using the actual progress toward validity. Think of it less like letting an artist freehand and more like drawing inside a stencil — you still get variation, but the lines cannot drift into impossible shapes. (github.com) ### Is validity the same as usefulness? No — and this is the catch. A chemically valid molecule can still be useless as a drug, hard to synthesize, toxic, unsta(nature.com) clears one bottleneck, but it does not solve target binding, ADMET, synthesis planning, or lab validation. The paper does try to go further by comparing 36 chemical-property distributions and by running a Turing-like expert plausibility test. (nature.com) ### Did it beat other models? The team says yes on standard benchmarks, and with much smaller models. The paper and project materials say(github.com)e GuacaMol benchmark while using up to an order of magnitude fewer parameters. That efficiency matters because molecule generation is only useful at scale if labs can actually run and adapt the models. (github.com) ### What did they build with it? They used the model to generate a database of 8.2 million synthetic molecules. That does not mean 8.2 million new drugs. But it does mean the team is pitching CoCoGraph as a pr(nature.com) giving chemists a much larger pile of plausible candidates to filter, rank, and test. (nature.com) ### So what changes now? The real shift is conceptual. A lot of generative chemistry work has accepted invalid outputs as the cost of exploration. CoCoGraph argues that you can keep the exploration and hard-wire away the dumbest failures. If tha(github.com)pelines, molecule AI gets a little less like autocomplete for chemistry and a little more like an actual design tool. (nature.com) ### Bottom line? This is not a drug-discovery miracle. But it is a meaningful engineering step — one that moves molecular generation from “make something interesting” toward “make something interesting that at least obeys the rules.” (nature.com)