Analyzing Atomic Interactions in Molecules as Learned by Neural Networks

2025/01/10 by Malte Esders, Thomas Schnake, Jonas Lederer +4 · 1 voice · 13 citations
Computer Science · Materials Science · #Artificial intelligence #Artificial neural network #Biological system #Computational Drug Discovery Methods #Computer science #Machine Learning in Materials Science #Machine learning #Molecule #Physics #Quantum #Quantum chemical #Quantum mechanics #Set (abstract data type) #Statistical physics #Topic Modeling

paper · doi:10.1021/acs.jctc.4c01424

published in Journal of Chemical Theory and Computation 21(2), 714-729 (American Chemical Society)

openalex publication_date 2025/01/10 · openalex created_date 2025/10/10 · openalex updated_date 2026/08/01

Abstract

While machine learning (ML) models have been able to achieve unprecedented accuracies across various prediction tasks in quantum chemistry, it is now apparent that accuracy on a test set alone is not a guarantee for robust chemical modeling such as stable molecular dynamics (MD). To go beyond accuracy, we use explainable artificial intelligence (XAI) techniques to develop a general analysis framework for atomic interactions and apply it to the SchNet and PaiNN neural network models. We compare these interactions with a set of fundamental chemical principles to understand how well the models have learned the underlying physicochemical concepts from the data. We focus on the strength of the interactions for different atomic species, how predictions for intensive and extensive quantum molecular properties are made, and analyze the decay and many-body nature of the interactions with interatomic distance. Models that deviate too far from known physical principles produce unstable MD trajectories, even when they have very high energy and force prediction accuracy. We also suggest further improvements to the ML architectures to better account for the polynomial decay of atomic interactions.

Citations

• Quantum-chemical insights from deep tensor neural networks
• SchNet – A deep learning architecture for molecules and materials
• Accurate computation of quantum excited states with neural networks

Cited by

• Training a force field for proteins and small molecules from scratch
• On the use of chemical bonding descriptors in machine learning
• Connecting the concepts of quantum state tomography and molecular representations for machine learning
• Data-driven complete basis set limit estimates from a minimal auxiliary basis
• Topological Feature Compression for Molecular Graph Neural Networks
• Implicit Machine Learning Force Fields Accelerate Molecular Dynamics Simulations
• Towards Desiderata-Driven Design of Visual Counterfactual Explainers
• Kernel Ridge Regression for conformer ensembles made easy with Structured Orthogonal Random Features
• Machine-Learned Extrapolation of Quantum Mechanical Energies in Implicit Solvent from Short to Long Oligopeptides. [europepmc]

Discussions

• ”Models that deviate too far from known physical principles produce unstable MD trajectories, even when they have very high energy and force prediction accuracy.” doi.org/10.1021/acs.... [bsky, 44 points, 2 comments]

Related

• Analyzing Atomic Interactions in Molecules as Learned by Neural Networks
• Analyzing atomic interactions in molecules as learned by neural networks
• Highly Accurate Description of Long-Range Interactions through the Combination of Neural Networks and Physical Models
• SchNet: A continuous-filter convolutional neural network for modeling quantum interactions
• Constructing accurate and efficient general-purpose atomistic machine learning model with transferable accuracy for quantum chemistry