Multimodal generation of astrocyte by integrating single-cell multi-omics data via deep learning

Abstract

ABSTRACTObtaining positive and negative samples to examining several multifaceted brain diseases in clinical trials face significant challenges. We propose an innovative approach known as Adaptive Conditional Graph Diffusion Convolution (ACGDC) model. This model is tailored for the fusion of single cell multi-omics data and the creation of novel samples. ACGDC customizes a new array of edge relationship categories to merge single cell sequencing data and pertinent meta-information gleaned from annotations. Afterward, it employs network node properties and neighborhood topological connections to reconstruct the relationship between edges and their properties among nodes. Ultimately, it generates novel single-cell samples via inverse sampling within the framework of conditional diffusion model. To evaluate the credibility of the single cell samples generated through the new sampling approach, we conducted a comprehensive assessment. This assessment included comparisons between the generated samples and real samples across several criteria, including sample distribution space, enrichment analyses (GO term, KEGG term), clustering, and cell subtype classification, thereby allowing us to rigorously validate the quality and reliability of the single-cell samples produced by our novel sample method. The outcomes of our study demonstrated the effectiveness of the proposed method in seamlessly integrating single-cell multi-omics data and generating innovative samples that closely mirrored both the spatial distribution and bioinformatic significance observed in real samples. Thus, we suggest that the generation of these reliable control samples by ACGDC holds substantial promise in advancing precision research on brain diseases. Additionally, it offers a valuable tool for classifying and identifying astrocyte subtypes.