Benchmarking pK_a Prediction Algorithms against an Extensive, Public Data Set.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Accurate prediction of proton dissociation constants (pKa) is vital for drug discovery and molecular modeling.
• Existing proprietary tools are popular, but open-source solutions are needed for large-scale computations.
• Machine learning (ML) offers a promising alternative to traditional empirical methods for pKa prediction.

Purpose of the Study:

• To address limitations in existing pKa prediction benchmarks, particularly regarding data set size and interpretation inconsistencies.
• To create a comprehensive, publicly accessible dataset of experimental aqueous pKa values.
• To benchmark the performance of various pKa prediction tools, including commercial and open-source ML models.

Main Methods:

• Assembled a large dataset of over 90,000 experimental aqueous pKa values for more than 31,000 unique molecules.
• Annotated the dataset with charge state transitions and microspecies distributions.
• Benchmarked seven pKa prediction methods: three commercial (ACD/Labs, Chemaxon, Epik) and four open-source ML models (MolGpKa, pKaSolver, QupKake, Uni-pKa).

Main Results:

• Developed the pKahub online database, one of the largest publicly available annotated pKa datasets.
• Provided a comprehensive benchmark of leading pKa prediction tools on a large, consistent dataset.
• Identified strengths and weaknesses of different pKa prediction approaches.

Conclusions:

• The pKahub dataset and benchmarking provide a valuable resource for advancing pKa prediction accuracy.
• Open-source ML models show promise as viable alternatives to commercial tools for pKa prediction.
• Improved pKa prediction is essential for efficient drug discovery and molecular modeling workflows.