Examining the effects of the aggregation function and alternative eco-friendly refrigerants using graph-based molecular machine learning

Abstract

Molecular structures that can be readily represented by graphs comprising constituent atoms (nodes) and their chemical bonds (edges) can also be used as input data for well-known machine learning (ML) models that process this data, such as graph neural networks (GNNs). GNNs have shown a reasonable performance in the predicting properties of chemical systems. In typical applications of GNNs to chemistry-related fields, the main objective is to create an optimal molecular representation by aggregating atomic features and pooling features in the graph. In this study, we investigated two different approaches that can possibly generate better molecular representations. First, we created intermolecular edges to predict the photochemical properties of chromophore molecules in the solution. These intermolecular edges were constructed using atomic partial charges, inspired from the fact that electrostatic interaction is the main component of solute-solvent interaction. In the second approach, we investigated the effect of the aggregation and pooling functions. The results showed that intermolecular electrostatic edges based on ground state charges prevent the GNN model from generating more effective molecular representations. On the contrary, the model demonstrated better performance when the averaging and adding operations were employed in a hybrid manner for the aggregation and pooling functions. Finding refrigerants is challenging because they should be non-toxic, non-flammable, energy-saving and thermodynamically stable. In addition, they should have low global warming potentials to mitigate global warming. Traditionally, a large number of molecules could be studied by quantum chemical calculations or experimental measurements. However, these methods are often challenging to apply on a large scale due to their time-consuming and resource-intensive nature. In this aspect, we propose a fully data-driven screening with machine learning models to find promising refrigerant candidates. We used two prediction methods: a less accurate but explicable feedforward neural network and a more accurate graph neural network (GNN). Using GNN, non-toxic molecules were determined using positive-unlabeled learning from the molecular dataset, which partially contains toxicity information. Out of 33k molecules, 7 molecules were recommended as future refrigerants. Molecular properties of these molecules and their uncertainties were also reported using data-driven methods. These contributed to the identification of alternative refrigerants, which remains a challenging topic today. We believe that our approach can be used to search for non-toxic emerging molecules in the future.