LAmbDA

LAmbDA implements a species- and dataset-independent transfer learning framework to map and label cell types in single-cell RNA sequencing (scRNA-seq) data while reducing batch effects and handling ambiguous labels.

Key Features:

• Species- and Dataset-Independence: Operates across scRNA-seq datasets from different species and experimental sources without requiring dataset-specific models.
• Transfer Learning Framework: Uses transfer learning to adapt models trained on one dataset for application to other datasets.
• Handling Ambiguous Labels and Batch Effects: Maps corresponding cell types between datasets to resolve ambiguous subtype labels while simultaneously reducing batch effects.
• High Accuracy in Subtype Labeling: Employs a Feedforward 1 Layer Neural Network with bagging and achieved weighted accuracies of 90% (simulated single-dataset), 94% (simulated two-dataset), 88% (pancreas datasets), and 66% (brain datasets).
• Superior Performance Compared to Other Methods: Outperformed scmap and CaSTLe in brain datasets (66% vs 60% for scmap and 32% for CaSTLe) and correctly predicted ambiguous subtype labels in 88% of test cases versus CaSTLe (63%), scmap (50%), and MetaNeighbor (50%).
• Generalizability Across Diverse Data Types: Applies across multiple model types and dataset compositions, supporting analysis of varied biological data.

Scientific Applications:

• Cross-species and Multi-dataset Cell Type Annotation: Annotating and mapping cell types across species and heterogeneous scRNA-seq datasets while mitigating batch effects and label ambiguity.
• Cellular Subtype Discovery: Identifying and labeling cellular subtypes in pancreas, brain, and simulated datasets to study subtype composition and function.
• Comparative Genomics and Data Integration: Integrating scRNA-seq data across experiments and species to support comparative analyses.

Methodology:

Applies transfer learning to map corresponding cell types between datasets, reduces batch effects, and uses a Feedforward 1 Layer Neural Network with bagging for subtype labeling.