Semisupervised Contrastive Learning for Bioactivity Prediction Using Cell Painting Image Data.

Area of Science:

• Computational Biology
• Machine Learning
• Drug Discovery

Background:

• Morphological profiling using Cell Painting images shows promise for predicting small molecule bioactivity.
• Existing machine learning methods for bioactivity prediction include fully supervised and self-supervised approaches.

Purpose of the Study:

• To introduce a novel semisupervised contrastive (SemiSupCon) learning approach for enhanced bioactivity prediction.
• To combine the benefits of supervised and self-supervised learning for analyzing Cell Painting data.

Main Methods:

• Developed and applied a semisupervised contrastive (SemiSupCon) learning framework.
• Utilized Cell Painting image data from two public datasets.
• Evaluated performance on MeSH pharmacological classifications and Drug Repurposing Hub annotations.

Main Results:

• SemiSupCon significantly improved downstream prediction performance for pharmacological classifications, mode of action, and biological targets.
• Successfully predicted biological activities of previously unannotated compounds, validated by literature.
• Demonstrated enhanced accuracy compared to existing methods.

Conclusions:

• The SemiSupCon approach effectively leverages both annotated and unannotated Cell Painting data for robust bioactivity prediction.
• This method has the potential to accelerate the discovery of biological activities from Cell Painting data with reduced human intervention.