DECAF: New AI method designs molecules by their 3D ensemble behavior

Most property-guided molecular design methods optimize for a single 3D conformer, ignoring that real molecules exist as ensembles (Boltzmann distributions). A new paper from Selma Moqvist et al. introduces DECAF (Decoupled Annealing Flows), which recasts molecular design as "Boltzmann-expected design." DECAF factorizes the joint distribution over molecular graphs and atomic coordinates into two conditional normalizing flows: a graph-conditioned flow p(x|G) that acts as a Boltzmann emulator (generating plausible conformers for a given graph), and a coordinate-conditioned flow p(G|x) that proposes new molecular graphs from a given 3D structure. By alternating these flows in a simulated-annealing loop, DECAF evaluates scoring functions on ensembles drawn from p(x|G), making ensemble statistics—not single-conformer properties—the target.

On the GEOM-Drugs benchmark, DECAF consistently shifted mean radius of gyration and solvent-accessible surface area toward desired targets, while single-conformer optimization degraded on larger drug-like molecules with broad Boltzmann distributions. Critically, DECAF requires no retraining to change objectives and extends to multi-objective trade-offs. It is the first 3D generative model capable of higher-moment design: jointly optimizing variance and skewness of an ensemble property to produce flexible molecules biased to a prescribed conformational regime—verified with all-atom MD simulations. This opens new avenues for designing molecules with tailored flexibility and dynamics.