GeoScatt-GNN: A geometric scattering transform-based graph neural network model for Ames mutagenicity prediction.

Abstract

This paper tackles the critical challenge of mutagenicity prediction by introducing three approaches. First, we demonstrate that 2D scattering coefficients derived from molecular images outperform traditional molecular descriptors. Second, we propose a hybrid framework unifying geometric graph scattering (GGS), Graph Isomorphism Networks (GIN), and machine learning models, achieving strong mutagenicity prediction. Third, we design, MOLG³-SAGE, a novel graph neural network architecture which integrates GGS node features into a fully connected graph structure, delivering state-of-the art predictive performance. Experimental results on the Hansen et al. dataset reveal that our methods surpass existing benchmarks, emphasizing the effectiveness of combining 2D and geometric scattering transform with graph neural networks. This study illustrates the potential of GNNs and GGS for mutagenicity prediction, with broad implications for drug discovery and chemical safety assessment.