SCGclust: Single Cell Graph clustering using graph autoencoders integrating SNVs and CNAs.

Area of Science:

• Genomics
• Computational Biology
• Cancer Research

Background:

• Intra-tumor heterogeneity (ITH) significantly impacts cancer prognosis and treatment efficacy.
• Single-cell DNA sequencing (scDNA-seq) offers cellular resolution for studying cancer progression and treatment responses.
• Accurate cell clustering is crucial for characterizing ITH from low-coverage scDNA-seq data.

Purpose of the Study:

• To develop a robust computational tool for cell clustering that integrates both single nucleotide variations (SNVs) and copy number alterations (CNAs).
• To improve the characterization of intra-tumor heterogeneity by leveraging complementary genomic signals.

Main Methods:

• A graph autoencoder and graph convolutional network (GCN) were co-trained to generate low-dimensional cell embeddings.
• A Gaussian Mixture Model was employed for subsequent cell clustering based on the embeddings.
• The method, SCGclust, was evaluated on simulated datasets and a real cancer sample.

Main Results:

• SCGclust demonstrated superior performance in cell clustering compared to existing SNV-based (SBMClone) and CNA-based (K-means) methods.
• The integration of both SNV and CNA signals led to more accurate characterization of ITH.
• Consistent improvements in V-measure scores were observed across multiple datasets.

Conclusions:

• Integrating SNV and CNA signals within a graph autoencoder framework enhances the accuracy of cell clustering for ITH analysis.
• SCGclust provides a powerful new approach for understanding cancer evolution and treatment resistance.
• The developed tool, SCGclust, is publicly available for the research community.