A pastry chef decorates a cake: they don't mold the whole thing at once but pipe the pattern step by step, twisting the nozzle at just the right angle. That's roughly how the new AI model called Atomic Design Transformer (ADT) works. It builds 3D molecules atom by atom, each time choosing which element to add and exactly where to attach it relative to the already placed "neighbors."
The trick is that ADT isn't tied to a global coordinate system. For each new atom, it uses a local reference frame from one of the previously placed atoms: this achieves invariance to rotations and translations of the whole molecule (a property physicists call SE(3)-invariance). The network itself is structured like a standard causal transformer, the architecture behind many language models, except here it predicts not the next word but the next atom.
Tests on the GEOM-Drugs dataset showed that a single trained model can handle different chemical "scaffolds"—from a simple benzene ring to six-membered heterocycles (pyridine, pyrimidine, pyrazine) and even furan with thiophene. XVR ranged from 11% for pyrazine to nearly 30% for benzene — modest but honest numbers for the task of generating drug-like molecules. At the same time, the quality of individual bonds and angles remained high even for larger molecules (up to 50 atoms). Generation is fast, and the diversity of structures is decent, making ADT a good "idea generator" for drug discovery.
The molecules in the GEOM-Drugs set are real compounds from databases, many of which are already used in pharmaceuticals. ADT learns from their shapes and can propose new, not-yet-synthesized variants.
A similar approach is described in Stanisław Lem's story 'Moloch': there, a computer assembled molecules according to specified properties until it created the perfect poison.