DRLM: A Robust Drug Representation Learning Method and its Applications.

Area of Science:

• Computational Biology
• Machine Learning
• Pharmacology

Background:

• High-quality drug representations are crucial for understanding pharmacology and advancing drug development.
• Existing methods for drug representation learning often overlook gene expression profiles, a valuable data source.
• Gene expression profiles from drug-treated cells contain cell-specific information that can hinder representation learning.

Purpose of the Study:

• To develop a novel deep learning method (DRLM) for learning robust drug representations from gene expression profiles.
• To integrate therapeutic use information into drug representations to enhance their predictive power.
• To mitigate the impact of cell specificity and noise in gene expression data.

Main Methods:

• A three-stage deep learning approach: stacked autoencoder for initial representation, iterative clustering to reduce cell specificity/noise, and a therapeutic use discriminator for information integration.
• Utilized gene expression profiles of drug-treated cells and drug therapeutic use information.
• Employed visualization analysis to assess representation quality and conducted extensive experiments on prediction tasks.

Main Results:

• DRLM effectively reduces cell specificity and noise in gene expression profiles.
• The learned drug representations incorporate therapeutic use information.
• DRLM-derived representations significantly outperform existing methods across various drug-related prediction tasks.

Conclusions:

• DRLM offers an effective strategy for learning drug representations by integrating gene expression profiles and therapeutic use data.
• The method successfully addresses challenges posed by cell specificity and noise.
• The improved drug representations facilitate more accurate predictions, aiding drug development.