CALLR

CALLR applies semi-supervised learning to annotate cell types in single-cell RNA sequencing (scRNA-seq) datasets, improving cell type label accuracy in complex tissues.

Key Features:

• Semi-supervised learning approach: Requires only a minimal number of labeled cell types to annotate the remaining cells.
• Integration of unsupervised and supervised learning: Combines unsupervised representation and supervised refinement to produce annotations.
• Graph Laplacian construction: Constructs a graph Laplacian matrix from all cells to represent the intrinsic data structure.
• Sparse logistic regression: Uses sparse logistic regression to refine cell type annotations based on available labeled data.
• Iterative optimization process: Alternates between updating cell clusters and annotation labels formulated as an optimization problem.
• Computational efficiency: Employs a computationally efficient algorithm to solve the posed optimization problem for large datasets.

Scientific Applications:

• Validation on real scRNA-seq datasets: Validated on 10 real scRNA-seq datasets to assess annotation accuracy.
• Benchmarking against existing methods: Demonstrated superior performance compared to existing (semi-)supervised learning methods and popular clustering techniques.

Methodology:

CALLR uses a semi-supervised framework that constructs a graph Laplacian from all cells, applies sparse logistic regression for supervised refinement, and alternates updates of clusters and annotation labels as an optimization problem solved by a computationally efficient algorithm.