Leveraging Cell Painting Images to Expand the Applicability Domain and Actively Improve Deep Learning Quantitative...

Area of Science:

• Drug Discovery
• Computational Chemistry
• Toxicology

Background:

• Drug discovery is lengthy and expensive.
• Ligand-based quantitative structure-activity relationship (QSAR) models optimize compound properties but have limited applicability domains.
• Image-informed models expand chemical diversity but require physically available compounds for imaging.

Purpose of the Study:

• To improve drug discovery by combining strengths of phenotypic and structure-based models.
• To enhance the performance of a mitochondrial toxicity assay (Glu/Gal).
• To expand the chemical space for reliable predictions.

Main Methods:

• Employed an active learning approach integrating phenotypic (Cell Painting) and structure-based models.
• Utilized a phenotypic Cell Painting screen to build a chemistry-independent model.
• Used model results to select compounds for experimental testing and annotation (Glu/Gal).

Main Results:

• Significantly improved the performance of a chemistry-informed ligand-based model.
• Successfully expanded the recognized chemical space by 10%.
• Demonstrated the efficacy of active learning in boosting model performance for toxicity prediction.

Conclusions:

• Active learning effectively combines phenotypic and structure-based approaches for enhanced drug discovery.
• The integrated method expands chemical diversity and improves prediction accuracy in toxicity assays.
• This strategy offers a more efficient and broader approach to identifying hit materials.