ResPAN

ResPAN: Deep Learning-Based Batch Correction for scRNA-seq Data

ResPAN performs batch effect correction and integration of single-cell RNA sequencing (scRNA-seq) datasets using a Wasserstein Generative Adversarial Network (WGAN) framework to reduce technical variation while preserving biological variability.

Key Features:

• WGAN Architecture: Implements a light-structured Wasserstein Generative Adversarial Network to model data distributions and correct batch effects while maintaining biological signals.
• Random Walk Mutual Nearest Neighbor Pairing: Aligns cells across batches by pairing cells with similar gene expression profiles to improve cross-dataset integration accuracy.
• Fully Skip-Connected Autoencoders: Incorporates skip-connected autoencoders to capture and retain critical biological information during batch correction.
• Large-Scale Scalability: Processes datasets exceeding 500,000 cells while maintaining integration performance.
• Benchmark Validation: Demonstrates superior batch correction and biological conservation compared to seven alternative methods using simulated and real scRNA-seq datasets.

Scientific Applications:

• Integrative scRNA-seq Analysis: Enables accurate identification of cell types, cell states, and lineage relationships across experimental conditions by minimizing batch effects.

Methodology:

ResPAN integrates scRNA-seq datasets by training a Wasserstein Generative Adversarial Network with random walk mutual nearest neighbor pairing to align cells across batches. Fully skip-connected autoencoders preserve biological information during adversarial training, producing corrected embeddings that reduce technical variability while conserving gene expression structure.