Implementation of multiple-instance learning in drug activity prediction.

Area of Science:

• Computational chemistry
• Cheminformatics
• Drug discovery

Background:

• Determining bioactive molecular conformers is crucial in drug discovery.
• Traditional methods face challenges in accurately identifying active structures.
• A novel approach using multiple-instance learning (MIL) is proposed.

Purpose of the Study:

• To implement MIL for predicting drug activity.
• To identify specific bioactive conformers for each molecule.
• To compare MIL with classical predictive models.

Main Methods:

• Molecules treated as bags of conformers; active if at least one conformer is active.
• Instance-based embedding using pharmacophore fingerprints and dissimilarity distances.
• Joint feature selection and classification using 1-norm Support Vector Machine (SVM).

Main Results:

• The MIL approach achieved top or second-best predictive models across four datasets.
• External validation confirmed its strong predictive abilities.
• 10 out of 12 known bioactive conformers were successfully identified.

Conclusions:

• The proposed MIL approach is powerful for drug activity prediction and avoids overfitting.
• It is highly competitive with classical predictive models.
• The method is validated as effective for identifying bioactive conformers.