Graph to Activation Energy Models Easily Reach Irreducible Errors but Show Limited Transferability

Abstract

Activation energy characterization of competing reactions is a costly but crucial step for understanding the kinetic relevance of distinct reaction pathways, product yields, and myriad other properties of reacting systems. The standard methodology for activation energy characterization has historically been a transition state search using the highest level of theory that can be afforded. However, recently, several groups have popularized the idea of predicting activation energies directly based on nothing more than the reactant and product graphs, a sufficiently complex neural network, and a broad enough data set. Here, we have revisited this task using the recently developed Reaction Graph Depth 1 (RGD1) transition state data set and several newly developed graph attention architectures. All of these new architectures achieve similar state-of-the-art results of ∼4 kcal/mol mean absolute error on withheld testing sets of reactions but poor performance on external testing sets composed of reactions with differing mechanisms, reaction molecularity, or reactant size distribution. Limited transferability is also shown to be shared by other contemporary graph to activation energy architectures through a series of case studies. We conclude that an array of standard graph architectures can already achieve results comparable to the irreducible error of available reaction data sets but that out-of-distribution performance remains poor.

3DReact: Geometric Deep Learning for Chemical Reactions

“…For the GDB7-22-TS set, however, all models result in unreasonable MAEs over 20 kcal/mol. This points to the particular challenges of the GDB7-22-TS set and suggests an avenue for further developments of ML models for extrapolative tasks …”

“…Atom-mapping provides static information, analogous to a reaction mechanism, to link atoms in reactants to atoms in products. While highly informative, and thought to be critical to the performance of 2D-graph-based models, − ,, accurate atom-maps are not available for all reaction data sets. ,, To circumvent the need for atom-mapping, but mimic its role in exchanging information between reactants and products, other approaches dynamically (i.e., in a learnable fashion) exchange information between molecular representations. For example, RXNMapper is a neural network that learns atom-mappings within the larger self-supervised task of predicting the randomly masked parts in a reaction sequence, using one head of a multihead transformer architecture.…”

“…Despite these advances for molecular property prediction, the prediction of computed reaction properties (principally, reaction barriers ,,− ) is still in its infancy . Machine learning approaches span from utilizing simple two-dimensional fingerprints of reaction components , (reactants and products) to physical-organic descriptors, ,,,,− or electronic structure-inspired features, to transformer models , adapted for regression, and 2D graph-based approaches. ,,, The latter, particularly the ChemProp model, , are often best-in-class in predicting reaction properties . It has been shown that these models achieve their impressive performance by exploiting atom-mapping information, − which provide information analogous to the reaction mechanism.…”

Virtual Special Issue on Machine Learning in Physical Chemistry Volume 2

“…In JPC A , the virtual special issue papers span applications to all aspects of chemical dynamics, molecular property prediction, and electronic structure. A large number of contributions within this collection in JPC A address fundamental research into new or adaptation of existing models for applications of ML to physical chemistry spanning many topical areas. − Many of the papers relate to using ML/AI and other data-driven models to enhance methods within physical chemistry. − A number of contributions address the creation or analysis of ground and excited state potential energy surfaces, − while others address dynamics, kinetics, and thermochemistry, a major area of interest within JPC A . − The use of ML to improve accuracy and efficiency in calculation of molecular properties is also addressed in many articles. − …”