Removing Biases from Molecular Representations via Information Maximization.

Area of Science:

• Biotechnology
• Computational Biology
• Drug Discovery

Background:

• High-throughput drug screening relies on cell imaging or gene expression data to link chemical structure to biological activity.
• Batch effects in large-scale screens introduce systematic errors, confounding non-biological associations and hindering accurate analysis.
• Accurate molecular representations are crucial for predicting drug properties and retrieving relevant molecules.

Approach:

• InfoCORE (Information maximization for COnfounder REmoval) is proposed to address batch effects in drug screening data.
• It employs variational lower bound on conditional mutual information to minimize batch-specific influences on latent representations.
• The method adaptively reweighs samples to equalize batch distributions, ensuring robust feature extraction.

Key Points:

• InfoCORE significantly improves molecular property prediction accuracy.
• It enhances molecule-phenotype retrieval performance in drug screening datasets.
• The approach demonstrates superior performance in mitigating batch effects compared to existing methods.

Conclusions:

• InfoCORE provides a versatile framework for removing batch effects and refining molecular representations.
• It effectively addresses general distribution shifts and data fairness issues by minimizing spurious correlations.
• The developed method offers a powerful tool for advancing drug discovery and biotechnology research.