Deep Learning Based Regression and Multiclass Models for Acute Oral Toxicity Prediction with Automatic Chemical...

Area of Science:

• Computational Chemistry
• Toxicology
• Machine Learning

Background:

• Acute oral toxicity (AOT) is crucial for compound safety assessment.
• In silico methods offer cost-effective alternatives to traditional toxicity testing.
• Existing predictive models require improvement in accuracy and efficiency.

Purpose of the Study:

• To develop novel deep learning models for accurate AOT prediction.
• To enhance the efficiency of toxicity assessment using molecular graph encoding.
• To explore the interpretability of deep learning models in toxicology.

Main Methods:

• Developed an improved molecular graph encoding convolutional neural networks (MGE-CNN) architecture.
• Constructed three high-quality AOT models: deepAOT-R (regression), deepAOT-C (multiclassification), and deepAOT-CR (multitask).
• Performed deep fingerprint exploration and reverse mining of features to identify AOT-related substructures.

Main Results:

• The deepAOT models demonstrated superior performance compared to previous methods on external datasets.
• deepAOT-R achieved R² of 0.864 and MAE of 0.195 on test set I.
• deepAOT-C and deepAOT-CR achieved prediction accuracies of over 95% on external test sets.
• Identified AOT-related chemical substructures through feature reverse mining.

Conclusions:

• The developed deep learning architecture provides a powerful and generalizable approach for predicting AOT and other chemical endpoints.
• The models offer significant improvements in accuracy and efficiency for in silico toxicity prediction.
• The freely available deepAOT models can aid researchers in compound safety evaluation.