Predicting compound activity from phenotypic profiles and chemical structures.

Area of Science:

• Computational chemistry
• Cheminformatics
• Pharmacology

Background:

• High-throughput screening (HTS) is crucial for drug discovery but resource-intensive.
• Virtual prediction of compound bioactivity can accelerate the identification of potential drug candidates.
• Integrating diverse data sources may improve predictive model performance.

Purpose of the Study:

• To evaluate the predictive power of chemical structures, Cell Painting (imaging), and L1000 (gene-expression) data for compound bioactivity.
• To determine the combined predictive strength of these three high-throughput data sources.
• To assess the potential of phenotypic profiling to enhance early-stage drug discovery.

Main Methods:

• Utilized a dataset of 16,170 compounds tested across 270 assays (585,439 readouts).
• Compared the predictive accuracy of chemical structures, Cell Painting, and L1000 profiles individually and in combination.
• Assessed the impact of combined data modalities on predicting compound activity across various assays.

Main Results:

• Each data modality (chemical structures, Cell Painting, L1000) individually predicted 6-10% of assays.
• Combining all three data sources achieved high accuracy in predicting 21% of assays, a 2-3 fold increase over single modalities.
• Using combined data increased predictable assays from 37% (chemical structures alone) to 64% (with phenotypic data).

Conclusions:

• Integrating chemical structures, imaging, and gene-expression data significantly enhances compound bioactivity prediction.
• Unbiased phenotypic profiling, particularly through Cell Painting, is a valuable tool for improving predictive models.
• This multi-modal approach accelerates early-stage drug discovery by increasing the efficiency of compound screening.