Interpretable Deep-Learning pK_a Prediction for Small Molecule Drugs via Atomic Sensitivity Analysis

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Machine learning models are vital for predicting drug properties like pKa, but often struggle with novel compounds due to limited data and lack of interpretability.
• Current models exhibit poor generalization to new molecules and lack transparency in their predictions.
• Understanding molecular ionizability is crucial for drug development and optimizing drug-target interactions.

Purpose of the Study:

• To develop a novel deep neural network (DNN) model, BCL-XpKa, for accurate prediction of drug acid-dissociation constants (pKa).
• To enhance model interpretability by enabling atomic-level insights into pKa predictions.
• To demonstrate the utility of BCL-XpKa and its associated atomic sensitivity analysis (ASA) in drug discovery workflows.

Main Methods:

• Developed BCL-XpKa, a DNN-based multitask classifier using Mol2D descriptors to encode local atomic environments for pKa prediction.
• BCL-XpKa outputs a discrete distribution, providing both pKa prediction and associated uncertainty.
• Implemented Atomic Sensitivity Analysis (ASA) to decompose pKa predictions into atomic contributions without retraining the model.

Main Results:

• BCL-XpKa demonstrates strong generalization to novel small molecules and competitive performance with existing ML pKa predictors.
• BCL-XpKa accurately models the impact of molecular modifications on ionizability.
• ASA successfully identified ionization sites in 93.2% of small molecule acids and 87.8% of bases, aiding in the optimization of a KRAS-degrading PROTAC.

Conclusions:

• BCL-XpKa offers a robust and interpretable solution for pKa prediction in drug development.
• Atomic sensitivity analysis provides valuable, high-resolution chemical insights from ML models.
• The combined approach of BCL-XpKa and ASA can effectively identify and address physicochemical liabilities in drug candidates.