SAMPL6 challenge results from predictions based on a general Gaussian process model.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Physical Chemistry

Background:

• Accurate prediction of a molecule's ionization state (pKa) is crucial for drug design and understanding physiochemical properties.
• The SAMPL6 blind challenge provided a benchmark for evaluating pKa prediction models using drug-like small molecules.

Purpose of the Study:

• To develop and evaluate a general Gaussian process regression model for predicting microscopic and macroscopic pKa values.
• To assess the model's performance on diverse, often polyprotic molecules in the SAMPL6 challenge.
• To incorporate uncertainty quantification into pKa predictions.

Main Methods:

• Utilized OpenEye Toolkits to generate molecular features related to proton removal.
• Employed Scikit-learn Gaussian process regression trained on 2700 macroscopic pKa values.
• Analytically determined macroscopic pKa values from predicted microscopic pKa values.

Main Results:

• The model achieved competitive performance in the SAMPL6 challenge, ranking among participants.
• Demonstrated promising agreement with experimental data despite training set limitations (predominantly monoprotic).
• The model's uncertainty estimates effectively indicated its domain of applicability and prediction accuracy.

Conclusions:

• The Gaussian process model offers a viable approach for pKa prediction with built-in uncertainty quantification.
• Improvements can be made by expanding the training set with more polyprotic molecules and refining ionizable group identification.
• The model's ability to predict its own accuracy is a valuable feature for practical applications.