scFANCL: Dual contrastive learning with false-negative correction at cell level for single-cell RNA-seq clustering

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell RNA sequencing (scRNA-seq) offers high-resolution cellular characterization but faces challenges in clustering due to high dimensionality, sparsity, and noise.
• Existing contrastive learning methods improve scRNA-seq representation but often overlook the inherent biological continuity within cell types.
• Dual contrastive frameworks capture cell-cell similarities and inter-cluster variations but typically focus on discrete cluster boundaries.

Purpose of the Study:

• To develop a novel dual contrastive framework, scFANCL, that effectively captures biological continuity in scRNA-seq data.
• To address the limitation of existing methods that neglect continuous transcriptional relationships among cells of the same type.
• To enhance the accuracy and biological relevance of scRNA-seq clustering.

Main Methods:

• Proposed scFANCL, a dual contrastive framework incorporating a cosine-similarity-based threshold to filter false negatives.
• Excluded cells of the same type from the negative pool to preserve continuous transcriptional relationships while maintaining inter-cluster separation.
• Evaluated scFANCL on seven public scRNA-seq datasets.

Main Results:

• scFANCL achieved competitive clustering performance, consistently yielding high Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI) scores across diverse datasets.
• Ablation studies confirmed the effectiveness of the false negative filtering component, demonstrating significant improvements over variants without filtering.
• Learned embeddings from scFANCL reflected biologically meaningful transcriptional transitions and differentiation trajectories.

Conclusions:

• scFANCL effectively preserves biological continuity within cell types while ensuring inter-cluster separation by excluding same-type cells from the negative pool.
• The framework demonstrates competitive clustering performance on benchmark scRNA-seq datasets.
• scFANCL's learned embeddings capture biologically relevant transcriptional structures and characteristics of rare cell populations.