Detecting retinal neural and stromal cell classes and ganglion cell subtypes based on transcriptome data with deep...

Area of Science:

• Genomics and Bioinformatics
• Computational Biology
• Ophthalmology Research

Background:

• Single-cell RNA sequencing (scRNA-seq) generates high-dimensional data for cell type identification.
• Accurate classification of retinal cell types and subtypes is crucial for understanding retinal function and disease.
• Existing methods may struggle with batch effects and require robust deep learning approaches.

Purpose of the Study:

• To develop and evaluate deep learning models for identifying diverse retinal cell types and retinal ganglion cell (RGC) subtypes.
• To assess model accuracy and robustness across multiple scRNA-seq datasets and experimental batches.
• To provide a validated computational tool for scRNA-seq data analysis in retinal research.

Main Methods:

• Development of deep domain adaptation models utilizing three distinct mouse scRNA-seq datasets.
• Application of four loss functions to align data distributions, minimize misclassification, and enhance model robustness.
• Evaluation of model performance using classification accuracy and confusion matrix analysis.

Main Results:

• The developed model achieved approximately 92% accuracy in classifying 39 retinal cell types from a large dataset.
• High accuracy (∼97%) was reached in classifying 40 and 45 RGC subtypes from two separate datasets.
• The lead model demonstrated accuracy ranging from 74% to nearly 100% across seven different data batches, indicating robustness.

Conclusions:

• Deep learning models can accurately and robustly identify retinal cell types and RGC subtypes from scRNA-seq data.
• The validated models are suitable for application to diverse scRNA-seq datasets, overcoming batch variations.
• The developed computational tools and datasets are publicly available to facilitate further research.