Deep learning in single-cell and spatial transcriptomics data analysis: advances and challenges from a data science...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Single-cell and spatial transcriptomics offer unprecedented insights into cellular functions and interactions.
• Analysis of high-dimensional, sparse, and multi-modal omics data presents significant challenges.
• Limited availability of high-quality annotated datasets and complex biological correlations hinder accurate reconstruction of cellular states.

Purpose of the Study:

• To systematically review advanced deep learning methods for analyzing single-cell and spatial transcriptomics data.
• To evaluate the performance of 58 computational methods across 21 benchmark datasets.
• To provide insights for selecting appropriate methods and identify future research directions.

Main Methods:

• Comprehensive literature review of deep learning techniques applied to transcriptomics.
• Curated 21 benchmark datasets from nine sources for method evaluation.
• Benchmarking of 58 computational methods using standardized metrics.

Main Results:

• Deep learning effectively addresses challenges posed by high-dimensional, sparse, and multi-modal omics data.
• Significant variation in method performance across different datasets and metrics was observed.
• The study provides a comparative analysis to guide the selection of computational approaches.

Conclusions:

• Deep learning holds significant potential for advancing transcriptomic data analysis in biological, medical, and clinical research.
• Understanding method performance variability is crucial for practical applications.
• Future development should focus on enhancing deep learning applications for complex biological data.