ML Potentials · Constructing Intermolecular Potentials by Combining Physics and Machine Learning
Horizon 2020 — Marie Skłodowska-Curie Actions
- Duration
- 2018-03-15 → 2020-03-14
- EU contribution
- €160,800
- Participants
- 1
- Scheme
- MSCA-IF
Lines connect the coordinator with its partners.
Results in brief
Constructing Intermolecular Potentials by Combining Physics and Machine Learning
Molecular interactions have a central role in determining the chemical and physical properties of molecules and materials. Intramolecular interactions (a.k.a. chemical bonds) bind atoms together to form molecules; on the other hand, intermolecular interactions are the attractive and repulsive forces between non-bonded fragments within a molecule or in-between different molecules. Even though intermolecular interactions contribute only a tiny fraction to the total energy, accurately quantifying intermolecular interactions is essential to correctly model chemical phenomena like the structure, dynamics, and function of proteins, the binding and metabolism of drugs, the structure and relative stability of supramolecular complexes and crystal polymorphs, and the orientation and reactivity of molecules on surfaces. Intermolecular potentials can be accurately quantified by applying high-level molecular quantum mechanics methods. Although these methods are physically sound and mathematically rigorous, their computational cost grows exponentially with the size of the system. Conversely, molecular mechanics compromises accuracy for speed, and approximates intermolecular interactions by parameterizing a classical potential energy function called a force field. Although molecular mechanics methods are applicable to systems with millions of atoms, they are unreliable for systems and physicochemical phenomena that are dissimilar to those used to parameterize the force field, which significantly limits their accurate applications. These limitations motivated me to develop state-of-the-art machine-learning (ML) models to accurately and rapidly predict molecular interaction energies and forces. Just as human chemists learn from past experiences to make predictions about the properties of new molecules, in ML a mathematical model is trained to leverage prior experimental and/or computational results to predict the properties of new molecules. This proposal is motivated by the impressive success of recent statistical ML models for accurately predicting molecular energies and forces, but their ineluctable shortcomings for modeling intermolecular (long-range and non-local) interactions and their consequent failure to scale to larger systems. By modeling long-range interactions using quantum mechanics methods (where they are computationally affordable), I developed an ML method that can be trained using only small-to-medium-size molecules, but applied to larger molecules. The main objectives achieved within the first half of the proposed action include: 1) developing a new methodology called Distance-Adapted version of SchNet (DASNet) which improve the existing state-of-the-art neural network architectures, 2) implementing a prototype of DASNet alongside designing the framework of the ultimate software package for predicting interaction energies based on the proposed model, 3) building the infrastructure for running quantum chemistry computations and compiling training data, 4) Releasing ChemTools software (free and open-source package which embodies a collection of interpretive tools for analyzing outputs of quantum chemistry calculations to gain chemical knowledge), 5) disseminating the action outcome by presenting at international workshops and conferences and organizing a hands-on workshop in Europe to promote Python programming language and teach ChemTools software package to a wide range of researchers.
Data: CORDIS, © European Union
Project objective
Statistical-learning approaches are emerging as powerful alternatives to direct approaches to solving the electronic Schrödinger equation for determining the energy and other properties of molecules. Despite the recent success of methods like deep neural networks, these methods are limited to relatively small molecules. The issue is that predicting long-range intermolecular interactions with machine learning requires sampling the vast diversity of chemical environments that occur on an extended length scale, leading to a combinatorial explosion in the amount of training data that is required. To solve this problem, the functional form of the long-range interactions is taken from physical models, but the parameters that enter those expressions (atomic charges/multipoles; induced charges/multipoles; van der Waals coefficients) are determined by combining physical insight with machine learning. In this model, machine learning is used only to predict short-range phenomena like the dependence of atomic charges/multipoles on the molecular structure and the dependence of induced atomic charges/multipoles on the local electric field. The resulting machine-learned physically-motivated atomistic intermolecular potentials are valid for molecules of any size, but only require training data from small- and medium-sized molecules.This development will provide molecular energies with the accuracy of quantum methods, at the computational cost of classical molecular mechanics approaches. This not only allows one to compute interaction energies for large molecules (e.g., the binding energy between a drug and a receptor), but allows the computational screening of molecules based on computed interaction energies. In addition to its transformative computational utility, this pioneering strategy—using physical insight to build a model, then using machine learning methods for the parameters in the model—can be extended to many other problems in chemistry, physics, and materials science.
Original text from CORDIS.
Participants
- UNIVERSITE DU LUXEMBOURG · ESCH-SUR-ALZETTECoordinatorLuxembourg
Links
Data: CORDIS, © European Union