ACTIVA

ACTIVA generates realistic synthetic single-cell RNA sequencing (scRNA-seq) data conditioned on cell-type information to augment analyses and improve reproducibility and classification of rare cell subpopulations.

Key Features:

• Single-stream adversarial variational autoencoder: Uses a single-stream adversarial variational autoencoder conditioned on cell-type information.
• Unified population and subpopulation generation: Produces synthetic data for whole populations or targeted subpopulations within a single model, avoiding separate models such as scGAN and cscGAN.
• Improved gene-level realism: Generates synthetic cells with improved pairwise correlations between genes, increasing realism relative to GAN-based models.
• Classifier indistinguishability: Synthetic cells are harder for classifiers to distinguish from real cells.
• Training on public scRNA-seq datasets: Trained on multiple public scRNA-seq datasets to support applicability across different contexts.
• Comparative performance: Demonstrates greater realism than scGAN and cscGAN and reduces training time by an order of magnitude compared to both scGAN and cscGAN.
• Data augmentation impact: Data augmentation with ACTIVA improves classification of rare subtypes by over 45% versus non-augmented data and by 4% versus cscGAN.

Scientific Applications:

• scRNA-seq pipeline augmentation: Augments scRNA-seq datasets to improve downstream analyses for smaller cohorts.
• Algorithm benchmarking: Provides synthetic data for benchmarking new computational methods on controlled, realistic data.
• Classifier evaluation: Enables assessment of classifier accuracy using realistic synthetic cells that challenge discrimination.
• Marker gene identification: Supports identification of marker genes with increased precision via augmented datasets.
• Rare subtype classification: Enhances classification performance for rare cell subtypes using targeted augmentation.
• Reduction of initial study size: Can reduce the number of patients and animals required in initial studies by improving analyses with fewer observations.

Methodology:

Employs a cell-type-conditioned single-stream adversarial variational autoencoder trained on multiple public scRNA-seq datasets and evaluated against scGAN and cscGAN using pairwise gene-correlation metrics, classifier-based indistinguishability, and training-time comparisons.