MultiGAI: Global Attention-Based Integration of Single-Cell Multi-Omics with Batch-Effects Correction

Abstract Single-cell multi-omics technologies allow the simultaneous measurement of multiple molecular modalities within the same cell, such as gene expression and chromatin accessibility (scRNA-seq + scATAC-seq) or gene expression and cell surface protein abundance (scRNA-seq + ADT), providing a multidimensional perspective on cellular states and regulatory mechanisms. However, these modalities often differ substantially in their distributions and noise levels and are affected by technical biases during experimental batches and sample processing, which can obscure true biological signals. To address these challenges, we present Multi-GAI, a variational autoencoder (VAE) framework with a global attention mechanism. MultiGAI integrates global information from the dataset during encoding and employs specially designed components to limit the propagation of batch information. This design allows effective batch effects correction while preserving key biological signals, generating high-quality latent representations of cells. Notably, in addition to performing well on single-cell multi-omics data, MultiGAI also demonstrates good performance in batch effects correction for single-cell transcriptomic data and in the integration of spatial transcriptomic data. Overall, MultiGAI provides a novel strategy for batch effects correction while retaining biological information, offering new insights for future single-cell multi-omics data integration.