Graph Representations and Neural Network Architectures for Reaction Barrier Height Prediction

Abstract

Chemical reactions are fundamental transformations that drive countless natural phenomena and technological applications, with their barrier heights, the minimum energy required for a reaction to potentially proceed, being crucial for understanding reactions. Deep learning approaches for predicting reaction barrier heights have shown promise but typically rely on explicit atom-to-atom mapping information, which is often unavailable in real-world scenarios. This thesis systematically explores graph representations and neural network architectures for predicting reaction barrier heights without relying on explicit atom mapping. We show that by incorporating reaction-specific inductive biases into neural network architectures, we can significantly reduce the performance gap between mapped and unmapped representations. Our best mapped-representation using Principal Neighbourhood Aggregation on the Condensed Graph of Reaction achieves a mean absolute error of 4.32 ± 0.45 kcal/mol, while our proposed Reaction Graph Transformer operating without atom mapping information reaches 6.18 ± 0.30 kcal/mol. The performance gap narrows even further on single reaction-type datasets, demonstrating that reaction-specific architectures can effectively compensate for missing mapping information. This work establishes a pathway toward more practical reaction property prediction tools that can operate effectively when atom mapping information is unavailable, enabling broader application in computational chemistry and drug discovery. Additionally, we contribute a flexible, modular code base for reaction property prediction that facilitates experimentation with different representations and architectures.