Decoding phenotypic screening: A comparative analysis of image representations.

Area of Science:

• Biomedical Imaging
• Computational Biology
• Machine Learning

Background:

• High content screening (HCS) is crucial for drug discovery but limited by high costs.
• The JUMP-CP consortium provides a large, open dataset for deep learning research in HCS.
• Developing universal representation models for HCS data is essential for broader accessibility.

Purpose of the Study:

• To develop a universal representation model for HCS data using the JUMP-CP dataset.
• To evaluate supervised and self-supervised learning approaches for HCS data representation.
• To assess model performance on mode of action and property prediction tasks.

Main Methods:

• Utilized the JUMP-CP dataset, focusing on U2OS cells and CellPainting protocol.
• Applied supervised and self-supervised deep learning techniques.
• Developed an evaluation protocol for phenotypic screening tasks.

Main Results:

• Self-supervised learning on multi-consortium partner data yielded representations robust to batch effects.
• The self-supervised approach achieved performance comparable to standard methods.
• The study identified effective training strategies for HCS image representation models.

Conclusions:

• Self-supervised learning offers a robust and effective method for HCS data representation.
• Leveraging diverse datasets enhances model generalizability and reduces batch effects.
• Recommendations are provided for optimizing representation model training in HCS.