The art of atom descriptor design

Area of Science:

• Computational Chemistry
• Quantum Mechanics
• Machine Learning
• Cheminformatics

Background:

• Traditional methods for understanding chemical phenomena rely on complex quantum mechanical (QM) calculations.
• Predicting properties like chemical reactivity, pKa values, and metabolic site requires detailed atomic information.
• The computational cost of QM calculations can be prohibitive for large-scale applications.

Purpose of the Study:

• To review various atom descriptor approaches for understanding chemical phenomena.
• To explore the substitution of QM calculations with machine learning (ML) models.
• To highlight the convergence of QM and ML into a new discipline, QM/ML.

Main Methods:

• Review of atom descriptors derived from quantum mechanics (wavefunctions, electron density) and classical descriptions.
• Analysis of ML models for predicting energies, forces, spectroscopic properties, and atomic charges.
• Focus on descriptors enabling fast calculation of atomic charges for force field parametrization.

Main Results:

• Atom descriptors provide insights into chemical reactivity, selectivity, pKa, metabolism prediction, and hydrogen bond strengths.
• ML models offer a viable alternative to QM calculations for various properties.
• Efficient calculation of atomic charges using descriptors facilitates force field development.

Conclusions:

• A deep understanding of chemical physics is crucial for designing effective atom descriptors.
• The integration of QM and ML is transforming computational chemistry.
• The emerging field of QM/ML promises faster and more accurate predictions in chemistry.