Capturing cell heterogeneity in representations of cell populations for image-based profiling using contrastive...

Area of Science:

• Computational biology
• Cellular imaging
• Machine learning

Background:

• Image-based cell profiling measures thousands of single-cell features to compare perturbed cell populations.
• Averaging single-cell data into sample profiles masks crucial cellular heterogeneity.
• Existing methods for aggregating cell profiles can be complex and may not fully capture population variations.

Purpose of the Study:

• To introduce CytoSummaryNet, a novel Deep Sets-based method for improved cell profile aggregation.
• To enhance mechanism of action (MoA) prediction accuracy in image-based cell profiling.
• To provide a more accessible and effective approach for analyzing single-cell feature data.

Main Methods:

• CytoSummaryNet utilizes a multiple-instance learning framework with self-supervised contrastive learning.
• The model aggregates single-cell features, moving beyond simple averaging.
• Training requires only readily available perturbation labels.

Main Results:

• CytoSummaryNet improved mechanism of action prediction by 30-68% in mean average precision compared to average profiling.
• Interpretability analysis revealed the model prioritizes larger, uncrowded cells while downweighting cells with debris or small mitotic cells.
• The method offers a straightforward post-processing step for single-cell profiles.

Conclusions:

• CytoSummaryNet effectively captures cellular heterogeneity, leading to significant improvements in MoA prediction.
• The self-supervised, multiple-instance learning framework provides an accessible yet powerful tool for cell profile aggregation.
• This approach offers a substantial advancement for retrieval performance in image-based profiling datasets.