source: arxiv statistics ml: boltzmann-expected molecular design with decoupled annealing flows
level: research
most 3d properties relevant to molecular design, like free energies and shape descriptors, are expectations over the boltzmann distribution of 3d configurations for a given molecular graph. existing property-guided generative models often tie each property to a single structure, ignoring the ensemble of possible conformations. this work recasts 3d molecular design as boltzmann-expected design, where the goal is to optimize properties averaged over the boltzmann ensemble.
the proposed method, decaf (decoupled annealing flows), factorizes the joint distribution over molecular graphs and 3d coordinates into two conditional flow models. one is a graph-conditioned flow that acts as a boltzmann emulator, generating 3d conformations for a given graph. the other is a coordinate-conditioned flow that proposes new graphs based on 3d information. by alternating between these two flows, decaf optimizes molecular graphs using a simulated annealing acceptance rule, where the scoring function is evaluated on ensembles drawn from the boltzmann emulator.
this approach allows decaf to directly optimize ensemble-averaged properties, leading to more realistic and physically meaningful molecular designs. the method is demonstrated on tasks where considering the full conformational ensemble is crucial, such as designing molecules with specific free energy profiles or shape distributions. decaf outperforms single-structure baselines by capturing the inherent flexibility and thermodynamic behavior of molecules.
why it matters: it enables ai-driven molecular design that accounts for realistic 3d flexibility, improving drug discovery and materials science.
source: arxiv statistics ml: boltzmann-expected molecular design with decoupled annealing flows