Benchmarking deep learning methods for biologically conserved single-cell integration.

scIB-E is a comprehensive deep learning-based benchmarking framework for evaluating single-cell RNA sequencing (scRNA-seq) data integration methods. Unified Benchmarking Framework: Evaluates 16 deep-learning single-cell integration methods using a unified variational autoencoder (VAE) framework. Incorporates batch information, cell-type labels, and combined strategies across three integration levels. Refined Metrics for Intra-cell-type Variation: Extends the single-cell integration benchmarking (scIB) metrics by adding new metrics to better capture intra-cell-type biological conservation. Novel Loss Function: Introduces Corr-MSE Loss, a correlation-based loss function designed to preserve global cellular relationships and enhance intra-cell-type biological variation. The preprocessed datasets are available at src/data.