ScribbleDom: using scribble-annotated histology images to identify domains in spatial transcriptomics data.

Area of Science:

• Spatial transcriptomics
• Computational biology
• Bioinformatics

Background:

• Spatial domain identification is crucial in spatial transcriptomics.
• Current unsupervised methods offer room for improvement.
• Lack of labeled data limits supervised approaches.

Purpose of the Study:

• To develop an enhanced spatial domain identification method.
• To leverage human expertise for improved accuracy.
• To introduce a semi-supervised learning approach.

Main Methods:

• Developed ScribbleDom, a semi-supervised convolutional neural network.
• Integrated human expert input (scribbles) for prior knowledge.
• Utilized a loss function combining gene expression similarity and expert annotations.
• Incorporated Inception blocks to capture spatial microenvironment information.

Main Results:

• ScribbleDom significantly improves spatial domain identification quality.
• Achieved substantial quantitative improvements (e.g., up to 169.38% adjusted Rand index).
• Demonstrated superior performance on benchmark datasets, including human brain and cancer samples.
• Maintained competitive results in fully unsupervised mode.

Conclusions:

• Semi-supervised learning with human input enhances spatial domain identification.
• ScribbleDom offers a powerful and flexible approach for spatial transcriptomics.
• The method provides a competitive alternative to existing unsupervised techniques.