JIND

JIND performs automated cell-type identification from single-cell RNA sequencing (scRNA-seq) data by using neural networks to learn a low-dimensional latent representation that enables robust annotation across datasets with batch effects.

Key Features:

• Low-Dimensional Latent Representation: Learns a low-dimensional latent code from scRNA-seq gene expression profiles that captures features required for reliable cell-type determination.
• Asymmetric Batch Alignment: Maps transcriptomic profiles of unseen cells onto a pre-established latent space via an asymmetric alignment strategy, avoiding model retraining for new datasets.
• Cell-Type-Specific Confidence Thresholds: Learns confidence thresholds specific to each cell type to identify and exclude cells that cannot be reliably classified.
• Performance and Efficiency: Demonstrates improved classification accuracy across datasets with batch effects while rejecting a small proportion of cells, with a parallelizable batch alignment reported to be over five to six times faster than Seurat integration.

Scientific Applications:

• Complex tissue composition analysis: Annotating cell types and resolving cellular composition in complex tissues and organisms.
• Developmental biology and differentiation studies: Analyzing differentiating cells and cellular differentiation processes.
• Disease modeling and precision medicine: Providing reliable cell-type annotations from scRNA-seq datasets to support disease modeling and precision medicine research.

Methodology:

Neural networks directly learn a low-dimensional latent code; an asymmetric alignment maps new dataset profiles into the pre-established latent space; cell-type-specific confidence thresholds are learned; the batch alignment step is parallelizable.