Leveraging DFT and Molecular Fragmentation for Chemically Accurate pK_a Prediction Using Machine Learning

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Physical Organic Chemistry

Background:

• Predicting molecular acidity (pKa) is crucial in chemistry.
• Existing methods often rely on computationally expensive calculations or lack accuracy.
• Density functional theory (DFT) calculations exhibit systematic errors for pKa prediction.

Purpose of the Study:

• To develop an accurate and efficient machine learning (ML) framework for predicting pKa values of complex organic molecules.
• To correct systematic errors in DFT calculations using ML.
• To explore the utility of a connectivity-based hierarchy (CBH) fragmentation protocol for generating molecular descriptors.

Main Methods:

• A random forest-based ML model was employed.
• Physics-based features from low-level DFT calculations were incorporated.
• Structural features were derived using the connectivity-based hierarchy (CBH) fragmentation protocol.
• The framework was validated on SAMPL6 and Novartis benchmark datasets.

Main Results:

• The ML framework accurately predicts pKa values of complex organic molecules.
• The combination of DFT and CBH features effectively corrects systematic DFT errors.
• The model demonstrates good generalizability and performance on benchmark datasets.
• Physics-based features reduce data dependence and the need for complex deep learning architectures.

Conclusions:

• The proposed quantum mechanical/ML framework offers an accurate and efficient approach for pKa prediction.
• The CBH fragmentation protocol is extended for generating novel molecular descriptors for ML applications.
• This method provides a valuable tool for computational chemists and drug discovery researchers.