Stacking Gaussian processes to improve predictions in the SAMPL7 challenge.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Accurate prediction of acid dissociation constants (pKa) is crucial for rational molecular design in pharmaceuticals.
• Existing machine learning methods, like Gaussian Process (GP) regression, offer fast pKa predictions but face limitations with diverse chemical spaces.
• Previous GP models struggled with accuracy due to insufficient chemical diversity in training data, particularly concerning polyprotic acids.

Purpose of the Study:

• To develop an advanced machine learning model for accurate pKa predictions.
• To address the limitations of standard GP models in handling broad chemical diversity and complex molecular structures.
• To improve the prediction of macroscopic pKa values for arbitrary chemical species.

Main Methods:

• Construction of a deep Gaussian Process (GP) model capable of incorporating more features without the curse of dimensionality.
• Training both standard GP and deep GP models on a curated database of approximately 3500 small molecules.
• Validation of the models on the SAMPL6 and SAMPL7 community-wide blind challenges.

Main Results:

• The deep GP model showed minor improvements over the standard GP for SAMPL6 predictions.
• The deep GP model demonstrated significant improvements for SAMPL7 macroscopic pKa predictions, achieving a Mean Absolute Error (MAE) of 1.5 pKa units.
• The enhanced model better handles datasets with limited ionizable sites or environments compared to the training set.

Conclusions:

• Deep Gaussian Process models offer a promising approach to enhance the accuracy of pKa predictions.
• The developed deep GP model shows superior performance in predicting macroscopic pKa values, especially for challenging chemical datasets.
• This advancement has implications for accelerating rational molecular design in the pharmaceutical industry and beyond.