Predicting Toxicity toward Nitrifiers by Attention-Enhanced Graph Neural Networks and Transfer Learning from Baseline Toxicity

Assessing chemical environmental impacts is critical
but challenging
due to the time-consuming nature of experimental testing. Graph neural
networks (GNNs) support superior prediction performance and mechanistic
interpretation of (eco-)toxicity data, but face the risk of overfitting
on the typically small experimental data sets. In contrast to purely
data-driven approaches, we propose a mechanism-guided transfer learning
strategy that is highly efficient and provides key insights into the
underlying drivers of (eco-)toxicity. By leveraging the mechanistic
link between baseline toxicity and toxicity toward nitrifiers, we
pretrained a GNN on lipophilicity data (log P) and subsequently fine-tuned
it on the limited data set of toxicity toward nitrifiers, achieving
prediction performance comparable with pretraining on much larger
but mechanistically less relevant data sets. Additionally, we enhanced
GNN interpretability by adjusting multihead attentions after convolutional
layers to identify key substructures, and quantified their contributions
using a Shapley Value method adapted for graph-structured data with
improved computational efficiency. The highlighted substructures aligned
well with and effectively distinguished known structural alerts for
baseline toxicity and specific modes of toxic action in nitrifiers.
The proposed strategy will allow uncovering new structural alerts
in other (eco)toxicity data, and thus foster new mechanistic insights
to support chemical risk assessment and safe-by-design principles.