High-Throughput Gene Expression Profiles to Define Drug Similarity and Predict Compound Activity

Area of Science:

• Pharmacology
• Computational Chemistry
• Genomics

Background:

• Exploring uncharted compound areas requires integrating biological information beyond chemical properties.
• Transcriptional data offers a novel dimension for compound analysis and similarity assessment.

Purpose of the Study:

• To integrate transcriptional data for 31,000 compounds using the L1000 platform.
• To assess the utility of transcriptional connection scores for compound similarity.
• To apply machine learning for target activity prediction and scaffold hopping analysis.

Main Methods:

• Utilized the L1000 platform to generate transcriptional profiles for 31,000 compounds.
• Developed and applied a transcriptional connection score, optimized using significant genes and confidence intervals.
• Employed Support Vector Machine models for predicting activity against NR3C1 and HSP90 targets.

Main Results:

• Optimized transcriptional connection score reduces noise and enhances reproducibility.
• Machine learning models achieved ≥80% balanced accuracy for NR3C1 and HSP90 activity prediction.
• Identified 22 novel HSP90-independent NR3C1 inhibitors with pyrimidine and pyrazolo-pyrimidine scaffolds.

Conclusions:

• Integrating transcriptional data significantly enhances compound discovery and target prediction.
• The developed methods enable the separation of specific and promiscuous compounds.
• The identified NR3C1 inhibitors offer potential therapeutic leads for depression treatment by targeting the glucocorticoid receptor without affecting its chaperone.