The virtual first principles based navigation of chemical space deepens our understanding of chemistry, and could help with the design of new experiments or materials. Unfortunately, the number of theoretically possible stable compounds is vast, readily exceeding the number of atoms in the solar system. Comprehensive screening therefore remains prohibitive in general, and the scope of sub-sampling efforts remains modest due to the computational cost of quantum chemistry methods such as density functional theory (DFT) or post-Hartree-Fock methods. Fortunately, quantum predictions throughout chemical space can be accelerated by multiple orders of magnitude with machine learning (ML). Quantum ML models are impressively efficient and accurate but have not yet acquired the necessary scalability and transferability to freely sample chemical space. Many ab initio fragmentation strategies have already addressed scalability but typically trade accuracy or transferability for speed, or suffer from steep scaling pre-factors. While transferable by design, popular cheminformatics models, e.g. based on extended connectivity fingerprint descriptors, have yet to be successfully applied to quantum properties. Here, we overcome this challenge through design and application of optimal localized training compound sets resulting in ML models with unprecedented efficiency, accuracy, scalability, and transferability. A sub-graph matching procedure systematically fragments any query compound into the smallest possible complete set of constituting effective quasi atoms in their chemical environment, dubbed Atoms-in-Molecules (am-ons) (A). Resulting AML models are scalable, efficient, and converge with respect to size (accuracy) and number (transferable) of amons. We numerically demonstrate robust performance, converging rapidly to low prediction errors on par with experimental uncertainty, for energies, NMR shifts, atomic charges and forces, core-electron excitation energies, and polarizability. Systems studied include the protein uibiquitin, chignolin, 11k organic molecules, drugs, DNA base-pairs, polymers, silicon, water clusters, and doped hBN sheets. Conceptually, amons encode the "DNA" of chemistry in terms of a dictionary of building blocks, effectively extending Mendeleev’s table to also account for chemical environments from which any query molecule can be constructed. The creation and use-case of such a dictionary signifies a dramatic dimensionality reduction of chemical space.

Towards efficient, accurate, scalable, and transferable quantum machine learning with AM-ons: The DNA of chemistry

2019 Spring National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.