Expanding Training Data for Structure-Based Receptor-Ligand Binding Affinity Regression through Imputation of Missing...

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Machine learning models require large datasets for effective training.
• Current structure-based molecular property prediction models face limited training data, particularly for receptor-ligand binding affinity.
• Existing datasets like CrossDocked2020 expanded binding pose classification data but not binding affinity data.

Purpose of the Study:

• To investigate the viability of imputing binding affinity labels to expand training data for structure-based molecular property prediction models.
• To assess the impact of using imputed labels on the performance of binding affinity regression models.

Main Methods:

• Utilized imputed binding affinity labels for molecular complexes lacking experimental data.
• Trained a convolutional neural network using existing binding affinity data from the CrossDocked2020 dataset.
• Evaluated the performance of structure-based models trained with both experimental and imputed binding affinity data.

Main Results:

• Imputing binding affinity labels is a feasible strategy for augmenting training datasets.
• Models trained with imputed labels demonstrated a marginal improvement in binding affinity regression performance.
• The inclusion of imputed labels introduced additional noise into the training data, yet performance still improved.

Conclusions:

• Imputation of binding affinity labels offers a practical method to increase training data for structure-based molecular modeling.
• Despite inherent noise, imputed data can enhance the predictive accuracy of machine learning models for receptor-ligand binding affinity.
• This approach contributes to advancing computational drug discovery by overcoming data limitations.