scDeepInsight: a supervised cell-type identification method for scRNA-seq data with deep learning.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Accurate cell-type annotation is crucial for analyzing single-cell RNA sequencing (scRNA-seq) data and understanding cellular heterogeneity.
• Current unsupervised clustering methods lack reference datasets, limiting cell-type classification accuracy and refinement.
• Supervised approaches are needed to improve the precision of cell-type identification in scRNA-seq.

Purpose of the Study:

• To introduce scDeepInsight, a novel supervised method for accurate cell-type annotation of scRNA-seq data.
• To enhance cell-type recognition by integrating manifold assignments, batch normalization, and outlier detection.
• To develop a method capable of identifying marker genes associated with specific cell types.

Main Methods:

• scDeepInsight converts tabular scRNA-seq data into images using the DeepInsight methodology.
• A trainable image transformer facilitates the conversion of non-image RNA data into image representations.
• Convolutional neural networks (e.g., EfficientNet-b3) process these images for automatic feature extraction and cell-type identification.

Main Results:

• scDeepInsight demonstrated superior performance compared to six other mainstream cell annotation methods.
• The method achieved an average accuracy rate of 87.5% in cell-type annotation.
• This represents a significant improvement of over 7% compared to existing state-of-the-art techniques.

Conclusions:

• scDeepInsight offers a powerful and accurate supervised approach for scRNA-seq cell-type annotation.
• The image-based deep learning strategy effectively addresses limitations of traditional clustering methods.
• This method advances the analysis of single-cell data, enabling more precise biological insights.