Self-supervised learning of cell type specificity from immunohistochemical images.

Area of Science:

• Bioimaging
• Proteomics
• Cell Biology

Background:

• In situ proteomic characterization of cell-cell interactions is crucial for understanding development and disease.
• Identifying cell-type-specific protein markers is essential for bioimaging, but current methods using single-cell transcriptomics have limitations due to gene-protein expression divergence.

Purpose of the Study:

• To develop a novel method for predicting cell type specificity of protein markers directly from unlabeled bioimaging data.
• To improve the accuracy and efficiency of identifying reliable protein markers for cell type classification in complex tissues.

Main Methods:

• A convolutional neural network was trained using a self-supervised objective to generate image embeddings.
• Non-linear dimensionality reduction was employed to visualize and analyze image clusters based on cell types and protein specificity.
• An image classifier was trained using cell type specificity estimates from independent single-cell transcriptomics data, without manual image labeling.

Main Results:

• The self-supervised model successfully clustered images according to cell types and anatomical regions.
• The developed method demonstrated superior classification of known proteomic markers in kidney tissue compared to marker selection via single-cell transcriptomics.
• The approach enables marker identification from unlabeled immunohistochemistry images.

Conclusions:

• This method offers a powerful way to identify cell type-specific protein markers from existing unlabeled bioimaging datasets.
• It overcomes limitations of gene-protein expression divergence and reduces the need for manual image annotation.
• The approach has significant implications for advancing bioimaging applications in biological research and disease diagnostics.