Enhancing Carbon Acid pK_a Prediction by Augmentation of Sparse Experimental Datasets with Accurate AIBL (QM) Derived...

Area of Science:

• Computational Chemistry
• Physical Organic Chemistry
• Cheminformatics

Background:

• Predicting aqueous pK_a for carbon acids is difficult due to a lack of high-quality experimental data.
• Quantitative Structure-Property Relationship (QSPR) and cheminformatics methods face challenges in developing accurate global models.

Purpose of the Study:

• To enhance the accuracy of aqueous pK_a prediction for carbon acids.
• To develop a computationally efficient method by integrating diverse data sources.

Main Methods:

• A novel method generating an atom-type feature vector, termed a distance spectrum, from the ionizable atom.
• Learning atom-type specific coefficients to quantify their impact on pK_a.
• Augmenting the dataset with pK_a values from high-performing local models, including the Ab Initio Bond Lengths (AIBL) method.

Main Results:

• The integrated model demonstrated reduced prediction error on an external test set compared to models using only literature experimental data.
• Distilling knowledge from multiple predictive models into a single general model improved overall performance.
• The distance spectrum approach effectively captures structural information relevant to pK_a.

Conclusions:

• Combining data from various sources, including computational models like AIBL, significantly improves the prediction of carbon acid aqueous pK_a.
• The developed method offers a more accurate and computationally efficient approach to pK_a prediction.
• This work contributes to advancing cheminformatics tools for chemical property prediction.