Specter

Specter performs scalable spectral clustering of large-scale single-cell RNA sequencing (scRNA-seq) and multi-modal data, including droplet-based sequencing technologies and CITE-seq antibody-derived tags (ADTs), to identify transcriptionally and proteomically distinct cell populations.

Key Features:

• Spectral clustering: Applies spectral clustering techniques to capture global data structure for downstream embedding and clustering.
• Linear-Time Complexity: Operates in linear time relative to the number of cells, demonstrated by clustering 2 million mouse embryo cells in 26 minutes.
• Landmark-based sparse representation: Uses landmarks to form a sparse representation of the full dataset, avoiding subsampling while retaining global relationships.
• Spectral embedding computation: Computes a spectral embedding from the landmark-based sparse representation in linear time.
• Cluster ensemble scheme: Aggregates multiple clusterings into an ensemble to improve clustering accuracy and robustness.
• Sensitivity to rare cell types: Enhances detection of subtle transcriptomic differences and low-abundance cell populations.
• Multi-modal data integration: Integrates RNA and protein measurements (e.g., gene expression and ADTs in CITE-seq) to resolve nuanced subpopulation differences.

Scientific Applications:

• Single-cell RNA-seq analysis: Identification of transcriptionally distinct cell groups within ultra-large scRNA-seq datasets to study cellular heterogeneity.
• Multi-modal omics studies: Joint analysis of gene expression and protein-marker (ADT) data from CITE-seq to dissect combined transcriptomic and proteomic cell states.

Methodology:

Adopts a landmark-based approach that avoids subsampling to build a sparse representation of the full dataset, computes a spectral embedding from that sparse representation in linear time, and combines results via a cluster ensemble scheme.