logD7.4 modeling using Bayesian Regularized Neural Networks. Assessment and correction of the errors of prediction

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Accurate prediction of the distribution coefficient (logD7.4) is crucial for drug development.
• Bayesian Regularized Neural Networks (BRNNs) offer a powerful framework for building predictive models.
• Automatic Relevance Determination (ARD) can enhance model interpretability and performance.

Purpose of the Study:

• To develop a predictive model for logD7.4 using BRNNs with ARD.
• To establish a method for assessing prediction accuracy based on compound similarity to the training set.
• To investigate the efficacy of dynamic correction of predictions using experimental data.

Main Methods:

• Utilized an in-house dataset of 5000 compounds with experimental logD7.4 values.
• Employed Bayesian Regularized Neural Networks (BRNNs) with Automatic Relevance Determination (ARD).
• Developed a distance-based metric to quantify prediction accuracy relative to the training data.
• Implemented dynamic correction of predictions using a continuously updated library of experimental logD7.4 values.

Main Results:

• A predictive model for logD7.4 was successfully constructed.
• A robust method for assessing prediction accuracy was established.
• Dynamic correction improved prediction reliability.
• Comparison showed local models and libraries provided no advantage over global models and libraries for homogeneous ionization classes.

Conclusions:

• BRNNs with ARD are effective for logD7.4 prediction.
• Distance-based accuracy assessment and dynamic correction enhance predictive models.
• Global models and libraries are sufficient for predicting logD7.4 across homogeneous ionization classes.