A machine learning-driven prediction of Hammett constants using quantum chemical and structural descriptors

Area of Science:

• Physical Organic Chemistry
• Computational Chemistry
• Machine Learning Applications

Background:

• The Hammett equation is crucial for modeling structure-activity relationships in physical organic chemistry.
• Predicting chemical reaction behavior and substituent effects remains a key challenge.
• Existing methods may not fully capture the complexity of substituent influences on reactivity.

Purpose of the Study:

• To apply machine learning (ML) for predicting Hammett constants (σm and σp) for benzoic acid derivatives.
• To develop and evaluate ML models using quantum chemical and molecular descriptors.
• To assess the predictive power and reliability of ML models for substituent effects.

Main Methods:

• Utilized a dataset of over 900 benzoic acid derivatives.
• Employed quantum chemical descriptors, Mordred-based electronic, steric, and topological descriptors.
• Trained and compared machine learning models, including Extra Trees (ET) and Artificial Neural Networks (ANNs).

Main Results:

• The Artificial Neural Network (ANN) model demonstrated superior performance with a test R² of 0.935 and RMSE of 0.084.
• The ANN model outperformed other tested ML models and a previously developed graph neural network.
• Feature importance analysis identified NBO charges and HOMO energies as key predictive descriptors.
• Applicability domain analysis confirmed the reliability and identified model limitations.

Conclusions:

• Machine learning, particularly ANNs, provides a robust and accurate method for predicting Hammett constants.
• This approach offers a valuable tool for chemical reactivity analysis and accelerated molecular design.
• The study underscores the potential of computational methods in advancing physical organic chemistry.