Anomaly detection for high-content image-based phenotypic cell profiling.

Area of Science:

• Cell Biology
• Computational Biology
• Bioinformatics

Background:

• High-content image-based phenotypic profiling uses automated microscopy to analyze cell morphology and infer physiological states.
• Traditional phenotypic profiles struggle to capture complex cell organization, and current machine learning methods lack biological interpretability.

Purpose of the Study:

• To develop a novel, interpretable representation for high-content phenotypic profiling.
• To improve the biological interpretability and downstream task performance of cell-based image analysis.

Main Methods:

• Utilized control well data to define the in-distribution of control experiments.
• Developed a self-supervised, reconstruction-based anomaly detection method to create interpretable representations.
• Evaluated anomaly-based representations against classical methods on four public Cell Painting datasets.

Main Results:

• Anomaly-based representations significantly improved reproducibility and Mechanism of Action classification accuracy.
• These representations effectively encoded complex morphological inter-feature dependencies while maintaining interpretability.
• Unsupervised explainability identified specific feature interactions driving observed anomalies.

Conclusions:

• Anomaly-based representations offer a powerful and interpretable approach for high-content phenotypic profiling.
• This method enhances downstream applications in cell biology, including drug discovery and mechanistic studies.
• The anomaly-based representation concept is adaptable to various cell biology image analysis challenges.