ScribbleDom

ScribbleDom applies a semi-supervised convolutional neural network to integrate expert scribble annotations and gene expression for identification of spatial domains in spatial transcriptomics.

Key Features:

• Semi-supervised convolutional neural network: Employs a semi-supervised CNN framework that leverages expert-provided annotations to guide spatial domain identification.
• Expert scribble integration: Integrates scribbles (annotations on histology images) as prior knowledge to influence domain assignments.
• Loss function combining expression and annotations: Uses a specialized loss function that combines similarity in gene expression profiles across tissue spots with adherence to human-provided annotations.
• Inception blocks for spatial continuity: Utilizes Inception blocks with convolution filters of varying sizes to extract microenvironmental information and account for spatial continuity.
• Optional unsupervised operation: Can operate in a fully unsupervised mode without human annotations while retaining competitive performance.

Scientific Applications:

• Spatial transcriptomics domain identification: Enhances identification of spatial domains by combining gene expression and expert annotations.
• Neuroscience (human dorsolateral prefrontal cortex): Demonstrated improvements in adjusted Rand index ranging from 1.82% to 169.38% for nine of twelve samples.
• Oncology (melanoma): Achieved a 15.54% improvement on a melanoma cancer dataset.

Methodology:

Computational methods explicitly include a semi-supervised convolutional neural network, integration of expert scribble annotations from histology images, a specialized loss function combining gene expression similarity and annotation adherence, Inception blocks with convolution filters of varying sizes to extract microenvironmental information, and an optional fully unsupervised mode.