INSCT

INSCT integrates single-cell RNA sequencing (scRNA-seq) data across heterogeneous batches, platforms, and species using batch-aware triplet neural networks to generate a unified embedding for downstream analysis.

Key Features:

• Batch Effect Mitigation: Learns a unified representation that minimizes distance between transcriptionally similar cells (Anchor and Positive) across batches while maximizing distance from dissimilar cells (Negative) within the same batch.
• Triplet Network Architecture: Feeds principal components of Anchor, Positive, and Negative into identical, weight-sharing neural networks trained with a triplet loss to produce an integrated two-dimensional embedding.
• Scalability and Efficiency: Demonstrated ability to integrate over 2.6 million transcriptomes from multiple independent studies in under 1.5 hours using less than 25 gigabytes of memory on a standard desktop configuration.
• Benchmark Performance: Comparative analyses indicate INSCT outperforms existing state-of-the-art scRNA-seq integration methods in scalability while maintaining comparable accuracy.
• Semi-Supervised Classification: Projects unlabeled cells into a reference collection of annotated cells to assign cell-type labels in independent scRNA-seq datasets.

Scientific Applications:

• Comprehensive Atlasing: Integration of large, heterogeneous scRNA-seq datasets to build unified cellular atlases.
• Developmental Biology: Alignment of cells across developmental timepoints and experimental batches to study lineage and differentiation.
• Neuroscience: Integration of neuronal and glial scRNA-seq datasets across platforms and studies to resolve cellular heterogeneity in the nervous system.
• Disease Modeling: Combining patient and control scRNA-seq datasets across batches and platforms to identify disease-associated cellular states.
• Large-Scale Integration Studies: Merging millions of transcriptomes from independent studies to enable cross-study comparative analyses.

Methodology:

Uses batch-aware triplet neural networks that take principal components of Anchor, Positive, and Negative cell profiles as input; identical weight-sharing networks are trained with a triplet loss to optimize inter- and intra-batch distances and produce a two-dimensional integrated embedding.