Impact of the characteristics of quantum chemical databases on machine learning predictions of tautomerization energies
arXiv:2104.06099 · doi:10.1021/acs.jctc.1c00363
Abstract
An essential aspect for adequate predictions of chemical properties by machine learning models is the database used for training them. However, studies that analyze how the content and structure of the databases used for training impact the prediction quality are scarce. In this work, we analyze and quantify the relationships learned by a machine learning model (Neural Network) trained on five different reference databases (QM9, PC9, ANI-1E, ANI-1 and ANI-1x) to predict tautomerization energies from molecules in Tautobase. For this, characteristics such as the number of heavy atoms in a molecule, number of atoms of a given element, bond composition, or initial geometry on the quality of the predictions are considered. The results indicate that training on a chemically diverse database is crucial for obtaining good results but also that conformational sampling can partly compensate for limited coverage of chemical diversity. We explicitly demonstrate that when certain types of bonds need to be covered in the target database (Tautobase) but are undersampled in the reference databases the resulting predictions are poor. A quantitative measure for these deficiencies is the Kullback-Leibler divergence between reference and target distributions. Analysis of the results with a TreeMAP algorithm provides deeper understanding of specific deficiencies in the reference data sets. Capitalizing on this information can be used to either improve existing databases or to generate new databases of sufficient diversity for a range of ML applications in chemistry.
arXiv comment42 pages, 10 figures of main manuscript plus 21 pages, 12 figures of supporting information
References in corpus (23)
- Gaussian Approximation Potentials: the accuracy of quantum mechanics, without the electrons
- Fast and Accurate Modeling of Molecular Atomization Energies with Machine Learning
- SchNet - a deep learning architecture for molecules and materials
- Interpretable machine learning: definitions, methods, and applications
- ANI-1: An extensible neural network potential with DFT accuracy at force field computational cost
- Machine Learning Force Fields
- PhysNet: A Neural Network for Predicting Energies, Forces, Dipole Moments and Partial Charges
- Explainable Artificial Intelligence: Understanding, Visualizing and Interpreting Deep Learning Models
- Machine learning for molecular simulation
- Less is more: sampling chemical space with active learning
- Machine learning prediction errors better than DFT accuracy
- Machine Learning of Molecular Electronic Properties in Chemical Compound Space
- Towards Explainable Artificial Intelligence
- A Fourth-Generation High-Dimensional Neural Network Potential with Accurate Electrostatics Including Non-local Charge Transfer
- ANI-1: A data set of 20M off-equilibrium DFT calculations for organic molecules
- Hierarchical modeling of molecular energies using a deep neural network
- Accurate molecular polarizabilities with coupled-cluster theory and machine learning
- Predicting molecular dipole moments by combining atomic partial charges and atomic dipoles
- Reactive Dynamics and Spectroscopy of Hydrogen Transfer from Neural Network-Based Reactive Potential Energy Surfaces
- High-Dimensional Potential Energy Surfaces for Molecular Simulations
- Machine learning the computational cost of quantum chemistry
- ML Models of Vibrating HCO: Comparing Reproducing Kernels, FCHL and PhysNet
- Isomerization and Decomposition Reactions of Acetaldehyde Relevant to Atmospheric Processes from Dynamics Simulations on Neural Network-Based Potential Energy Surfaces