Bayesian optimization in chemical compound sub-spaces using low-dimensional molecular descriptors

Abstract Efficient optimization of molecules with targeted properties remains challenging due to the vast size and discrete nature of chemical compound space. Existing optimization approaches often rely on high-dimensional molecular representations and large training datasets, limiting their applicability in small-data regimes. We present a data-efficient optimization framework that combines three components: (i) a very low-dimensional, physics-informed molecular descriptor that substantially reduces the dimensionality of the optimization problem, (ii) Bayesian optimization for efficient exploration of the resulting descriptor space, and (iii) a modular inverse mapping framework converting optimized descriptor vectors into chemically valid molecular structure. Together, these components provide a bridge between continuous optimization and discrete chemical space. We benchmark the framework on the QM9 and PubChemQC B3LYP/6-31G*//PM6 datasets by optimizing molecular entropy, zero-point vibrational energy, normalized electronic energy, and HOMO–LUMO gap. The framework identifies molecules with target properties using as few as hundreds of molecular evaluations in most cases, and at most a few thousands evaluations, across chemical subspaces containing tens of thousands to millions of candidate molecules. It achieves success rates of up to 100 % for several optimization tasks. These results demonstrate that low-dimensional, physics-based molecular descriptors enable accurate surrogate modeling and data-efficient optimization, providing a transferable framework for molecular discovery in discrete chemical spaces.