Interpreting Neural Network Models for Toxicity Prediction by Extracting Learned Chemical Features

Area of Science:

• Computational chemistry
• Toxicology
• Machine learning

Background:

• Neural networks are widely used for chemical toxicity prediction.
• Their complex nature hinders interpretability, limiting user confidence in predictions.
• Existing interpretation methods like SHAP focus on input features but not internal network transformations.

Purpose of the Study:

• To develop a novel technique for interpreting neural network predictions in chemical toxicology.
• To identify specific chemical substructures responsible for activating hidden neurons within the network.
• To explain model predictions by linking hidden neuron activity to learned chemical representations.

Main Methods:

• A new interpretation method was developed to identify chemical substructures linked to hidden neuron activation.
• The importance of hidden neurons was assessed for individual test compounds.
• Substructures associated with activated neurons were used to explain predictions.
• The approach was validated against known mutagenicity structural alerts from the Derek Nexus system.

Main Results:

• The novel technique successfully identified chemical substructures responsible for hidden neuron activation.
• Explanations derived from this method were found to be competitive with established feature attribution techniques.
• The approach provides complementary insights into how neural networks learn and predict toxicity.

Conclusions:

• The developed method offers a powerful new way to interpret complex neural network models in toxicology.
• It enhances transparency by revealing the chemical basis for model predictions.
• This technique can improve confidence and trust in machine learning-based toxicity assessments.