Hydration free energies from kernel-based machine learning: Compound-database bias.

Area of Science:

• Computational Chemistry
• Machine Learning
• Physical Chemistry

Background:

• Accurate prediction of thermodynamic properties like hydration free energies is essential for drug discovery and materials science.
• Current computational methods often struggle with large chemical spaces and require efficient predictive models.

Purpose of the Study:

• To develop and evaluate a kernel-based machine learning approach for predicting hydration free energies of small organic molecules.
• To investigate the impact of training data diversity and potential biases on model performance and transferability.

Main Methods:

• Utilized atomistic simulations with implicit solvent models for calculating hydration free energies.
• Developed a kernel-based machine learning model incorporating conformational averaging and an atomic-decomposition ansatz.
• Employed dimensionality reduction and cross-learning techniques to analyze model learning rates and biases.

Main Results:

• The machine learning model demonstrated improved transferability due to the atomic-decomposition approach.
• Significant biases were identified in experimental compound databases, impacting model accuracy.
• The rate of learning was highly dependent on the breadth and variety of the training dataset.

Conclusions:

• Kernel-based machine learning, with appropriate feature representation and diverse data, can accurately predict hydration free energies.
• Fitting models to narrow chemical datasets leads to severe biases and limits predictive power.
• Future efforts should focus on curating diverse datasets to enhance the generalizability of machine learning models in chemistry.