scSMD: a deep learning method for accurate clustering of single cells based on auto-encoder

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) provides critical insights into cellular heterogeneity, development, and disease.
• Advancements in scRNA-seq necessitate sophisticated computational methods for data analysis.
• Managing sparse, high-dimensional scRNA-seq data is a key challenge.

Purpose of the Study:

• To apply deep learning techniques for improved single-cell data clustering.
• To develop a model capable of handling sparse and high-dimensional transcriptomic data.
• To enhance the analysis of cellular complexity and disease mechanisms using scRNA-seq.

Main Methods:

• Development of the SMD deep learning model, incorporating nonlinear dimensionality reduction.
• Utilizing a convolutional autoencoder architecture informed by the negative binomial distribution.
• Integration of a porous dilated attention gate for dynamic feature weighting.

Main Results:

• The SMD model demonstrates efficacy in precise single-cell data clustering.
• Successful application on both public and proprietary osteosarcoma datasets.
• Efficient capture of essential cell clustering features and dynamic adjustment of feature weights.

Conclusions:

• Deep learning, specifically the SMD model, shows significant potential for advancing scRNA-seq data analysis.
• The SMD model offers a robust framework for dissecting cellular complexities and understanding biological processes.
• This approach can enhance the elucidation of disease mechanisms through advanced transcriptomic analysis.