scNym

scNym predicts cell types and learns representations from single-cell RNA sequencing (scRNA-seq) data using semi-supervised adversarial neural networks to enable annotation transfer across experiments.

Key Features:

• Semi-Supervised Learning: Leverages both labeled and unlabeled datasets to learn rich representations of cell identities when comprehensive annotations are not available.
• Adversarial Neural Network Architecture: Uses an adversarial component to encourage the model to produce outputs consistent with real data distributions, improving generalization across datasets.
• Representation Learning: Derives meaningful low-dimensional representations from single-cell profiling data to support classification and downstream analyses.
• Annotation Transfer Across Experiments: Transfers cell identity annotations between distinct experiments despite biological and technical variation.
• Integration Across Datasets: Synthesizes information from multiple training and target datasets to adapt to new experimental conditions.
• Calibration and Interpretability: Produces well-calibrated predicted probabilities and supports interpretability via saliency methods that identify features driving predictions.
• Superior Performance: Demonstrates improved accuracy in transferring annotations across datasets compared to existing annotation methods.

Scientific Applications:

• Cell type classification and annotation: Assigns cell identities in scRNA-seq experiments to facilitate characterization of cellular composition.
• Comparative studies of cellular heterogeneity: Enables analysis of changes in cell populations across biological states or treatments.
• Cross-experiment annotation transfer: Supports transferring annotations between experiments to harmonize labels across datasets.

Methodology:

The model is trained on labeled data while simultaneously using unlabeled data in a semi-supervised framework; an adversarial objective enforces consistency with real data distributions, and information is integrated from diverse datasets to adapt to new experimental conditions and improve annotation accuracy.