Toxigraphnet: a graph neural network framework for precise toxicity prediction of drug molecules

Accurate prediction of a drug molecule's toxicity is a critical step in pharmaceutical research, offering the potential to reduce experimental costs, mitigate adverse effects, and accelerate drug development. Traditional computational methods often rely on handcrafted molecular descriptors, which fall short in capturing the intricate structural and chemical nuances of molecules. In this study, we propose ToxiGraphNet, a graph neural network (GNN)-based regression model for predicting the LD50 value-a quantitative measure of acute toxicity-directly from molecular SMILES strings. Using RDKit, molecules are transformed into graph representations where atoms serve as nodes and bonds as edges, each enriched with chemically meaningful features. Atom features encompass atomic type, degree, aromaticity, chirality, and more, while bond features capture bond type, conjugation, and ring status. These molecular graphs are processed via edge-conditioned convolution layers (NNConv) within the PyTorch Geometric framework, enabling dynamic, chemistry-aware feature aggregation. The model architecture includes three NNConv layers with batch normalization, dropout, and a residual connection to ensure stable training. After global mean pooling, the learned graph-level representations are passed through fully connected layers to predict LD50 values. Training on a curated LD50 dataset yielded impressive performance (MSE: 0.3610, MAE: 0.4424, RMSE: 6009, [Formula: see text]: 0.5959), demonstrating strong generalization and predictive accuracy. This work highlights the efficacy of GNNs in modeling molecular toxicity without relying on hand-engineered features and presents a scalable solution for property prediction in drug discovery pipelines.