In silico prediction of pKa values using explainable deep learning methods

Negative logarithm of the acid dissociation constant (p<i>K</i> _a) significantly influences the absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties of molecules and is a crucial indicator in drug research. Given the rapid and accurate characteristics of computational methods, their role in predicting drug properties is increasingly important. Although many p<i>K</i> _a prediction models currently exist, they often focus on enhancing model precision while neglecting interpretability. In this study, we present GraFp<i>K</i> _a, a p<i>K</i> _a prediction model using graph neural networks (GNNs) and molecular fingerprints. The results show that our acidic and basic models achieved mean absolute errors (MAEs) of 0.621 and 0.402, respectively, on the test set, demonstrating good predictive performance. Notably, to improve interpretability, GraFp<i>K</i> _a also incorporates Integrated Gradients (IGs), providing a clearer visual description of the atoms significantly affecting the p<i>K</i> _a values. The high reliability and interpretability of GraFp<i>K</i> _a ensure accurate p<i>K</i> _a predictions while also facilitating a deeper understanding of the relationship between molecular structure and p<i>K</i> _a values, making it a valuable tool in the field of p<i>K</i> _a prediction.