scTACL: a multitask topology-aware contrastive learning approach for single-cell transcriptomics analysis.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) enables gene expression profiling at the single-cell level, revealing cellular heterogeneity.
• scRNA-seq data often suffers from noise and dropout events, hindering downstream analysis.

Purpose of the Study:

• To develop a robust computational method for improving the quality and interpretability of scRNA-seq data.
• To address challenges posed by noise and dropout events in scRNA-seq datasets.

Main Methods:

• Introduced scTACL, a novel method utilizing contrastive learning between cell similarity and embedding similarity graphs.
• Employed a zero-inflated negative binomial (ZINB) distribution for data reconstruction.
• Integrated cell similarity graph and cell embedding similarity graph alignment.

Main Results:

• scTACL demonstrated superior performance in data imputation, clustering, batch effect correction, and cell-cell interaction analysis.
• Successfully identified two distinct epithelial cell subtypes in lung adenocarcinoma.
• Accurately distinguished epithelial and mesenchymal regions in liver cancer spatial transcriptome data without spatial information.
• Identified the COLLAGEN signaling pathway's role in epithelial-mesenchymal transition via intercellular communication analysis.

Conclusions:

• scTACL effectively enhances scRNA-seq data analysis by mitigating noise and dropout events.
• The method provides valuable insights into cellular heterogeneity and intercellular communication.
• scTACL is a versatile tool for complex biological applications, including cancer subtyping and spatial transcriptomics analysis.