DeepChemStable: Chemical Stability Prediction with an Attention-Based Graph Convolution Network.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Chemical instability in stored compounds can compromise drug discovery bioassay results.
• Accurate prediction of chemical stability is crucial but complex, often relying on empirical data models.
• Existing models using COMDECOM data have limitations in predicting chemical instability.

Purpose of the Study:

• To develop an advanced model for predicting chemical instability in drug compounds.
• To improve the accuracy and reliability of chemical stability predictions, particularly reducing false negatives.
• To introduce an end-to-end deep learning approach for chemical instability prediction.

Main Methods:

• Utilized the COMDECOM (COMpound DECOMposition) dataset for empirical data.
• Developed DeepChemStable, an attention-based graph convolution network model.
• Employed an end-to-end learning approach that dynamically learns structural features.

Main Results:

• DeepChemStable achieved an AUC value of 84.7%.
• The model demonstrated a recall rate of 79.8% for instability prediction.
• Achieved 79.1% accuracy through 10-fold stratified cross-validation.

Conclusions:

• DeepChemStable effectively predicts chemical instability, outperforming previous methods.
• The model's ability to dynamically learn features and reduce false negatives marks a significant advancement.
• This approach enhances the reliability of drug discovery processes by improving compound stability assessment.