scGAD

scGAD performs generalized cell type annotation and discovery for single-cell RNA-seq (scRNA-seq) data by labeling target cells with known cell types or novel cluster identities to enable identification of both recognized and novel cell types.

Key Features:

• End-to-End Algorithmic Framework: Integrates annotation of known cell types and discovery of novel ones within a single framework.
• Intrinsic Correspondence Building: Identifies geometrically and semantically mutual nearest neighbors as anchor pairs to establish correspondences between seen and novel cell types.
• Soft Anchor-Based Self-Supervised Learning Module: Leverages similarity affinity scores to transfer label information from reference datasets to target data while aggregating novel semantic knowledge in the prediction space.
• Confidential Prototype Self-Supervised Learning Paradigm: Employs a confidential prototype self-supervised learning paradigm to improve inter-type separation and intra-type compactness and to capture global topological structure in the embedding space.
• Bidirectional Dual Alignment Mechanism: Aligns both the embedding space and prediction space bidirectionally to mitigate batch effects and cell type shifts.
• Extensive Validation and Benchmarking: Demonstrated on large simulation datasets and real-world data with marker gene identification used to validate the biological significance of novel cell types.

Scientific Applications:

• Developmental Biology: Enables precise cell type classification and discovery in developmental biology datasets.
• Immunology: Supports identification and annotation of immune cell subtypes.
• Cancer Research: Facilitates cell type classification and discovery in cancer research.
• Downstream Analyses: Supports differential expression studies, pathway analysis, and biomarker identification linked to annotated and discovered cell types.

Methodology:

Identifies geometrically and semantically mutual nearest neighbors as anchor pairs; computes similarity affinity scores for soft anchor-based self-supervised label transfer and aggregation of novel semantics; applies a confidential prototype self-supervised learning paradigm to enhance inter-type separation and intra-type compactness and capture global topological structure; performs bidirectional dual alignment of embedding and prediction spaces to mitigate batch effects and cell type shifts.