PredPS

PredPS predicts compound stability in human plasma using an attention-based graph neural network, classifying compounds as stable (≥ 85% remaining at 3 hours) or unstable (< 85% remaining at 3 hours) to support early drug discovery and development.

Key Features:

• Model architecture: An attention-based graph neural network is used for prediction.
• Classification threshold: Compounds are classified as stable if ≥ 85% remains at 3 hours and unstable if < 85% remains at 3 hours in human plasma.
• Training data: The model leverages both in-house and open-source datasets.
• Performance metrics: Reported performance is AUROC 90.1%, overall accuracy 83.5%, sensitivity 82.3%, and specificity 84.6%.
• Evaluation protocol: Performance was assessed using 5-fold cross-validation.
• Comparative performance: The approach outperforms previously used traditional machine learning and deep learning algorithms for the same task.

Scientific Applications:

• Early drug discovery screening: Predicts plasma stability to inform selection of compounds for further development.
• Compound prioritization for lead optimization: Supports prioritization of compounds to improve systemic drug exposure and potential in vivo efficacy.
• Drug development decision support: Provides stability predictions in human plasma to guide downstream experimental testing and development choices.

Methodology:

Model training used an attention-based graph neural network trained on in-house and open-source datasets and evaluated by 5-fold cross-validation.