In silico prediction of pK _a values using explainable deep learning methods

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• The acid dissociation constant (pKa) is critical for predicting drug absorption, distribution, metabolism, excretion, and toxicity (ADMET).
• Computational methods offer rapid and accurate pKa prediction, essential for modern drug research.
• Existing pKa models often prioritize accuracy over interpretability, limiting insights into structure-property relationships.

Purpose of the Study:

• To develop an interpretable pKa prediction model using graph neural networks (GNNs) and molecular fingerprints.
• To enhance the understanding of how molecular structure influences pKa values.
• To provide a reliable computational tool for drug discovery.

Main Methods:

• Utilized graph neural networks (GNNs) combined with molecular fingerprints for pKa prediction.
• Developed separate models for acidic and basic pKa predictions.
• Integrated Integrated Gradients (IGs) to provide visual interpretability of atomic contributions to pKa.

Main Results:

• Achieved mean absolute errors (MAEs) of 0.621 for acidic and 0.402 for basic pKa predictions on the test set.
• Demonstrated good predictive performance for both acidic and basic compounds.
• Successfully visualized the atoms significantly influencing predicted pKa values using Integrated Gradients.

Conclusions:

• GraFpKa offers a reliable and interpretable approach to pKa prediction.
• The model enhances understanding of the relationship between molecular structure and pKa.
• GraFpKa serves as a valuable tool for accelerating drug discovery and development.