Leveraging bounded datapoints to classify molecular potency improvements.

Area of Science:

• Computational chemistry and cheminformatics
• Machine learning in drug discovery
• Quantitative structure-activity relationship (QSAR) studies

Background:

• Molecular machine learning (ML) models excel at predicting drug candidate potency.
• Traditional regression algorithms struggle with bounded inhibition data (e.g., IC50 values).
• This limitation hinders lead optimization and molecular discovery processes.

Purpose of the Study:

• To develop a novel computational approach for handling bounded inhibition data in drug discovery.
• To introduce a molecular pairing strategy that transforms regression problems into classification tasks.
• To evaluate the performance of this new approach against traditional methods.

Main Methods:

• Developed a molecular pairing approach to create a classification task: predicting which of two molecules is more potent.
• Utilized established ML algorithms like XGBoost and Chemprop for the new classification task (DeltaClassifiers).
• Validated the approach across 230 ChEMBL IC50 datasets, comparing performance with traditional regression models.

Main Results:

• Both tree-based and neural network-based DeltaClassifiers demonstrated improved accuracy in classifying molecular potency gains over regression methods.
• The Chemprop-based deep DeltaClassifier showed superior performance for paired molecules, regardless of scaffold similarity.
• This highlights the effectiveness of the classification approach for molecular optimization and scaffold-hopping.

Conclusions:

• The molecular pairing classification approach effectively addresses limitations of traditional regression for bounded inhibition data.
• DeltaClassifiers offer a promising alternative for enhancing molecular potency prediction and guiding drug discovery pipelines.
• This method shows significant potential for accelerating the identification of potent drug candidates and enabling scaffold-hopping strategies.