Resolution limit of data-driven coarse-grained models spanning chemical space

Abstract

Increasing the efficiency of materials design and discovery remains a significant challenge, especially given the prohibitively large size of chemical compound space. The use of a chemically transferable coarse-grained model enables different molecular fragments to map to the same bead type, while also reducing computational expense. These properties further increase screening efficiency, as many compounds are screened through the use of a single coarse-grained simulation, effectively reducing the size of chemical compound space. Here, we propose new criteria for the rational design of coarse-grained models that allows for the optimization of their chemical transferability and evaluate the Martini model within this framework. We further investigate the scope of this chemical transferability by parameterizing three Martini-like models, in which the number of bead types ranges from five to sixteen for the different force fields. We then implement a Bayesian approach to determining which chemical groups are more likely to be present on fragments corresponding to specific bead types for each model. We demonstrate that a level of performance and accuracy comparable to Martini can be obtained by using a force field with fewer bead types. However, the advantage of including more bead types is a reduction of uncertainty with respect to back-mapping these bead types to specific chemistries. Just as reducing the size of the coarse-grained particles leads to a finer mapping of conformational space, increasing the number of bead types yields a finer mapping of chemical compound space. Finally, we note that, due to the relatively large size of the chemical fragments that map to a single martini bead, a clear resolution limit arises when using the water/octanol partition free energy as the only descriptor when coarse-graining chemical compound space.

References in corpus (11)

• Global Optimization by Basin-Hopping and the Lowest Energy Structures of Lennard-Jones Clusters Containing up to 110 Atoms
• Two-dimensional materials from high-throughput computational exfoliation of experimentally known compounds
• Interatomic potentials from first-principles calculations: the force-matching method
• Big Data of Materials Science - Critical Role of the Descriptor
• Big Data meets Quantum Chemistry Approximations: The $\Delta$-Machine Learning Approach
• Alchemical and structural distribution based representation for improved QML
• Non-covalent interactions across organic and biological subsets of chemical space: Physics-based potentials parametrized from machine learning
• Feature Optimization for Atomistic Machine Learning Yields A Data-Driven Construction of the Periodic Table of the Elements
• In silico screening of drug-membrane thermodynamics reveals linear relations between bulk partitioning and the potential of mean force
• Efficient potential of mean force calculation from multiscale simulations: solute insertion in a lipid membrane
• Controlled exploration of chemical space by machine learning of coarse-grained representations

Cited by in corpus (5)

• Physics-inspired structural representations for molecules and materials
• Machine Learning of Free Energies in Chemical Compound Space Using Ensemble Representations: Reaching Experimental Uncertainty for Solvation
• Data-driven equation for drug-membrane permeability across drugs and membranes
• Condensed-phase molecular representation to link structure and thermodynamics in molecular dynamics
• Broad chemical transferability in structure-based coarse-graining