Multimodal-based crystal graph convolution neural networks for predicting soil toxicity to earthworms

Modeling approaches have been developed to quantitatively assess chemical toxicity in soil. However, an integrated modeling framework that incorporates molecular-level features, exposure conditions in soil, and the intrinsic properties of both organisms and chemicals remains underexplored. Therefore, this study aimed to develop a multimodal deep learning model that predicts toxicity indices by integrating multi-scale features across different domains. Micro-level features of chemicals were extracted using a crystal graph convolutional neural network, which is designed to learn structural representations from atomic and bonding information. These micro-level features were combined with macro-level data such as exposure, soil and organism-related conditions for predictions. The model was evaluated by predicting lethal concentration values causing 50 % mortality (LC50) of 14 chemicals for earthworms, the standard soil test species. For validation, its performance was compared with single-modal models using only micro or macro-level features. The late fusion-based multimodal model yielded a coefficient of determination reaching 0.86, and an average of 0.67 across cross-validation, despite the limited size of literature-derived dataset. This performance outperformed benchmark models. In addition, the model provided interpretable insights by identifying key input features for predicting LC50. Topological polar surface area and ionization energy emerged as the dominant macro- and micro-level contributors, respectively. These results highlight the potential of the proposed modeling approach for soil toxicity prediction and mechanistic understanding. Moreover, this study suggests the extensibility of a new modeling framework that integrates heterogeneous data for interpretable chemical toxicity assessment.