Drug activity prediction using multiple-instance learning via joint instance and feature selection.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Identifying bioactive molecular conformers and representative features is crucial in drug discovery.
• Experimental determination of bioactive conformers is challenging, necessitating computational methods like machine learning.
• High-dimensional feature spaces in Multiple Instance Learning (MIL) can hinder performance due to irrelevant or redundant features.

Purpose of the Study:

• To develop a novel computational approach for simultaneous instance and feature selection in Multiple Instance Learning (MIL).
• To address the challenge of high dimensionality in feature spaces for improved classification and interpretability in drug discovery.

Main Methods:

• Proposed a novel approach: multiple instance learning via joint instance and feature selection (MIL-JIFS).
• Employed iterative joint instance and feature selection using instance-based feature mapping.
• Utilized 1-norm regularized optimization for the selection process.

Main Results:

• The proposed MIL-JIFS approach was evaluated on four biological activity datasets.
• Selected instances (prototype conformers) and features (pharmacophore fingerprints) demonstrated competitive discriminative power.
• The selection process exhibited fast convergence.

Conclusions:

• The novel MIL-JIFS method effectively identifies key molecular conformers and features for biological activity prediction.
• The approach offers a computationally efficient and interpretable solution for drug discovery challenges.
• Empirical results validate the discriminative power and speed of the proposed method.