Data-Driven Modelling of Electronic Properties in Organic Radicals Using Graph Neural Networks

Abstract

Organic radicals play a central role in many areas of chemistry, but predicting their electronic properties through quantum chemical calculations remains computationally demanding. This project explores the use of artificial intelligence to address this challenge, focusing on properties related to the stability and electronic structure of organic radicals. To evaluate this strategy, a large-scale dataset containing approximately 246,000 radical and molecules was analysed using both classical molecular descriptors and graph-based representations. Classical molecular representations based on Morgan fingerprints and SOAP descriptors were compared with self-supervised graph representations obtained through the Molecular Contrastive Learning (MolCLR) framework. Different neural network architectures, including local propagation and attention-based models, were also evaluated. The results showed that graph-based representations outperform traditional molecular descriptors for the prediction of electronic properties in organic radicals. MolCLR provided a more effective description of the local chemical environment than classical 2D and 3D descriptors. In addition, hyperparameter optimisation further improved model performance. The resulting model predicted quantum chemical properties at a significantly lower computational cost than DFT calculations. Finally, latent-space analysis and clustering revealed that chemically related radicals are organised into coherent structural families within the learned representation. These findings indicate that graph neural networks not only improve predictive accuracy but also capture chemically relevant information about radical structures. Overall, this work demonstrates that graph-based machine learning methods can provide accurate and computationally efficient alternatives to traditional quantum chemical calculations. It also opens the possibility of predicting EPR-related descriptors directly from molecular structure in future studies.