scProfiterole: Clustering of Single-Cell Proteomic Data Using Graph Contrastive Learning via Spectral Filters.

Area of Science:

• Proteomics
• Computational Biology
• Bioinformatics

Background:

• Single-cell proteomics (scProteomics) data analysis tools lag behind single-cell gene expression (scRNAseq) due to data challenges like drop-outs and noise.
• Graph contrastive learning (GCL) shows promise for cell type identification in scProteomics, but requires robust handling of noisy similarity graphs.

Purpose of the Study:

• To develop a computational framework, scProfiterole, for effective clustering of scProteomics data using spectral graph filters within GCL.
• To address limitations in applying spectral filters to noisy scProteomics graphs by using polynomial interpolation via Arnoldi orthonormalization.

Main Methods:

• Introduced scProfiterole, a framework utilizing spectral graph filters (random walks, heat kernels, beta kernels) for GCL-based clustering.
• Implemented polynomial interpolation of spectral filters using Arnoldi orthonormalization to overcome computational limitations of direct filter application.
• Evaluated performance on comprehensive scProteomics datasets.

Main Results:

• scProfiterole with GCL and learnable polynomial coefficients improved cell type identification effectiveness and robustness.
• Heat and beta kernels enhanced clustering performance compared to adjacency matrices or random walks.
• Polynomial interpolation of spectral filters outperformed approximation or truncation methods.

Conclusions:

• scProfiterole provides an effective computational framework for scProteomics data clustering and cell type identification.
• The use of spectral graph filters, particularly heat and beta kernels, combined with polynomial interpolation significantly improves GCL performance on noisy scProteomics data.