Evaluating Classical and Quantum Graph Convolutional Architectures for Molecular Property Prediction

Abstract

Graph-based learning methods are an approach to molecular property prediction, where chemical structures are naturally represented as graphs. This work presents a simple test of three classical models which include Graph Convolutional Networks (GCN), Simplified Graph Convolution (SGC), and Linear Graph Convolution (LGC). These models were evaluated on chemical datasets and analyzed model behavior. These analyses highlight differences in graph-based learning architectures, providing insight into how model simplifications influence numerical behavior. Motivated by this perspective, we discussed quantum analogues of these models. The quantum models were formally considered in this thesis, while their empirical evaluation was not performed. The present study concentrates on classical results and establishes a common methodological framework for subsequent classical–quantum comparisons in molecular graph learning.