QSAR Modeling Based on Conformation Ensembles Using a Multi-Instance Learning Approach.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Drug Discovery

Background:

• Quantitative Structure-Activity Relationship (QSAR) models are crucial for designing bioactive molecules in drug discovery.
• Traditional QSAR models using 2D descriptors miss spatial structure information, while 3D QSAR models struggle with selecting the correct bioactive conformation.
• The multi-instance (MI) learning approach offers a potential solution by incorporating multiple molecular conformations into model training.

Purpose of the Study:

• To implement and evaluate various multi-instance learning algorithms for QSAR modeling.
• To compare the performance of MI-QSAR models against classical single-instance QSAR (SI-QSAR) approaches.
• To assess the ability of MI algorithms to automatically identify plausible bioactive conformations.

Main Methods:

• Implementation of conventional and deep learning-based multi-instance algorithms.
• Comparison of MI-QSAR models with SI-QSAR models using 2D descriptors and single lowest energy 3D conformations.
• Model performance evaluation across 175 datasets from the ChEMBL23 database.

Main Results:

• MI-QSAR models demonstrated superior predictive performance compared to SI-QSAR models in a significant number of cases.
• The study confirmed that MI algorithms can effectively identify relevant bioactive conformations without prior selection.
• This indicates a more robust and accurate approach to QSAR modeling, especially for complex molecular structures.

Conclusions:

• Multi-instance QSAR represents a significant advancement over single-instance QSAR for drug discovery applications.
• The MI approach mitigates the challenge of selecting optimal conformations, leading to improved model reliability.
• MI-QSAR facilitates the design of potentially bioactive molecules by leveraging comprehensive conformational information.