Compound Activity Prediction with Dose-Dependent Transcriptomic Profiles and Deep Learning.

Area of Science:

• Computational biology
• Drug discovery
• Systems pharmacology

Background:

• Predicting compound activity across diverse assays is crucial in drug discovery.
• Gene expression signatures from profiling assays can predict compound activity in other assays.
• Applications include predicting mechanism-of-action (MoA), off-target effects, and polypharmacology.

Purpose of the Study:

• To introduce transcriptomics-to-activity transformer (TAT) models for predicting compound activity in biochemical and cellular assays.
• To leverage gene expression profiles across multiple compound concentrations for enhanced prediction accuracy.

Main Methods:

• Developed TAT models using gene expression data from RASL-seq assays.
• Trained models to predict the activity of 2692 compounds across 262 dose-response assays.
• Validated model utility using a held-out dataset and prospective experimental testing.

Main Results:

• Achieved useful predictive models for 51% of the tested assays.
• Prospectively validated TAT predictions in a malaria inhibition assay, achieving a 63% hit rate.
• Identified several submicromolar malaria inhibitors through prospective validation.

Conclusions:

• Transcriptomic responses across compound concentrations, modeled by TAT, offer a powerful predictive framework.
• TAT models provide a cost-efficient method for identifying compound bioactivities in various assays.
• This approach holds significant potential for accelerating drug discovery and development.