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

Area of Science:

• Environmental Science
• Computational Toxicology
• Cheminformatics

Background:

• Quantitative assessment of soil chemical toxicity is crucial for environmental protection.
• Existing models often lack integration of multi-scale features (molecular, exposure, organismal).
• An underexplored area is a unified framework combining diverse data for toxicity prediction.

Purpose of the Study:

• To develop a multimodal deep learning model for predicting soil chemical toxicity indices.
• To integrate micro-level chemical features with macro-level exposure and organism data.
• To enhance the interpretability of toxicity prediction models.

Main Methods:

• Utilized a crystal graph convolutional neural network (CGCNN) for micro-level chemical feature extraction.
• Combined CGCNN features with macro-level data (exposure, soil, organism conditions).
• Employed a late fusion strategy within a multimodal deep learning framework.

Main Results:

• The multimodal model achieved a coefficient of determination of 0.86 for predicting earthworm LC50 values.
• Performance surpassed single-modal benchmark models, demonstrating the benefit of data integration.
• Identified key predictive features: topological polar surface area (macro) and ionization energy (micro).

Conclusions:

• The proposed multimodal deep learning approach shows significant potential for accurate soil toxicity prediction.
• The model offers mechanistic insights by highlighting critical contributing factors to toxicity.
• This framework is extensible for integrated, interpretable chemical risk assessment using heterogeneous data.