scAMAC: self-supervised clustering of scRNA-seq data based on adaptive multi-scale autoencoder.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) is crucial for understanding cellular heterogeneity and biological processes.
• Deep learning methods are increasingly used for scRNA-seq data clustering, but often lose structural information.
• Existing models struggle to capture inter-layer network interactions, limiting analytical power.

Purpose of the Study:

• To develop a novel self-supervised deep learning method for scRNA-seq data analysis.
• To improve cell clustering and data reconstruction by preserving network structural information.
• To offer a robust tool for downstream analyses like cell trajectory inference.

Main Methods:

• Developed scAMAC, a self-supervised clustering method using an adaptive multi-scale autoencoder.
• Employed a Multi-Scale Attention mechanism to fuse features across autoencoder layers (encoder, hidden, decoder).
• Utilized an adaptive feedback mechanism for parameter updates and optimized clustering via a membership matrix.

Main Results:

• scAMAC effectively fuses multi-scale features and preserves network structure for improved clustering.
• Demonstrated superior performance over advanced methods in both data clustering and reconstruction tasks.
• Showcased utility in downstream analyses, including cell trajectory inference.

Conclusions:

• scAMAC provides a powerful new approach for scRNA-seq data analysis, enhancing clustering and reconstruction.
• The method's ability to capture multi-scale cellular correlations and structural information is key to its success.
• scAMAC offers a valuable tool for advancing biological discovery from scRNA-seq data.