pK(a) prediction from "Quantum Chemical Topology" descriptors

Area of Science:

• Computational Chemistry
• Medicinal Chemistry
• Quantitative Structure-Property Relationships (QSPR)

Background:

• Accurate prediction of pK(a) is crucial for pharmaceutical development.
• Existing pK(a) prediction methods have limitations.
• Quantum Chemical Topology (QCT) offers novel ab initio descriptors.

Purpose of the Study:

• To develop and validate accurate pK(a) prediction models for carboxylic acids.
• To explore the utility of QCT descriptors in QSPR studies.
• To compare different computational methods and machine learning algorithms.

Main Methods:

• Utilized QCT to generate ab initio molecular descriptors.
• Compared five levels of theory, focusing on HF/6-31G(d) and B3LYP/6-311+G(2d,p).
• Developed predictive models using Partial Least Square (PLS), Support Vector Machines (SVM), and Radial Basis Function Neural Networks (RBFNN), validated with cross-validation.

Main Results:

• HF/6-31G(d) with SVM yielded the best predictive performance (q(2) = 0.886, RMSE = 0.293 for all acids).
• High accuracy was achieved across different subsets: ortho- (q(2)=0.825), para-/meta- (q(2)=0.923), and aliphatic (q(2)=0.906).
• The developed method demonstrated competitive performance against commercial pK(a) prediction software.

Conclusions:

• QCT-derived descriptors combined with SVM provide a powerful and accurate approach for pK(a) prediction.
• The HF/6-31G(d) level of theory offers a good balance between accuracy and computational cost.
• This methodology holds significant potential for accelerating drug discovery and development.