Predicting complexation performance between cyclodextrins and guest molecules by integrated machine learning and...

Area of Science:

• Pharmaceutical Science
• Computational Chemistry
• Drug Delivery

Background:

• Pharmaceutical formulation development is complex, often requiring extensive experimentation.
• Machine learning (ML) shows promise in advancing scientific prediction and analysis.
• Predictive models for drug/cyclodextrin (CD) systems are crucial for formulation design.

Purpose of the Study:

• To develop high-accuracy predictive models for drug/cyclodextrin complexation using machine learning.
• To identify key molecular descriptors influencing drug-CD interactions.
• To compare the predictive performance of different ML algorithms and molecular modeling.

Main Methods:

• Utilized molecular descriptors and experimental conditions as input features for ML models.
• Employed light gradient boosting machine (LGBM), random forest, and deep learning algorithms.
• Calculated complexation free energy as the output prediction.
• Performed molecular simulation for structural, dynamic, and energetic insights.

Main Results:

• The LGBM model achieved high accuracy, with a mean absolute error of 1.38 kJ/mol and R-squared of 0.86.
• LGBM outperformed random forest and deep learning models in predictive performance.
• Analysis identified critical molecular descriptors affecting drug-CD interactions.
• In ketoprofen-CD systems, ML models surpassed traditional molecular modeling predictions.

Conclusions:

• Developed ML models accurately and rapidly predict the solubilizing capacity of cyclodextrin systems.
• The integration of ML and molecular simulation offers synergistic benefits for formulation prediction.
• This research represents a significant advancement in applying ML to pharmaceutical formulation design.