MipSScs: Artificial neural network-based data integration of 2D/3D single-cell spatial RNA sequence data from...

There is interest in the use of recent single-cell spatial transcriptomic technologies to gain biological insights into disease mechanisms. Previously, we characterized the use of herpes simplex virus 1 (HSV-1) induced neuroinflammation in 2D dissociated cells from human cerebral organoids (dcOrgs) to model molecular and transcriptomic readouts associated with Alzheimer's disease (AD). In this work, we generated two datasets by using single-cell non-spatial RNA sequencing and single-cell spatial RNA sequencing technologies on HSV-1-infected 2D dcOrgs and HSV-1-infected 3D cerebral organoids (cOrgs). We conducted cell type assignment for the cells in the 2D dcOrgs and 3D cOrgs, by using single-cell non-spatial RNA sequence data from human fetal brains and adult post-mortem brains, to infer the transcriptomic effects of AD-associated in-vitro perturbations through viral infections linked to AD. We evaluated computational and machine learning methods, including the use of multi-layer perceptrons (MLPs), and we used cross-2D/3D platform comparisons as a benchmark to evaluate the artificial neural network models. In the process, we found that the use of MLPs can lead to high validation rates for assigning cell type identities from 2D and 3D human cerebral organoids to cell types found in human adult post-mortem brain samples. Furthermore, the use of these technologies and systems enabled the identification of pseudotime trajectories and cell clusters associated with the viral transcriptional life cycle. We identified several cell types, including endothelial cells and astrocytes, with significantly more clustered cell-cell nearest neighbor distances in infected 3D cOrgs compared to mock 3D cOrgs. Permutation tests revealed that these differences in nearest neighbor distances are unlikely to be driven by overall structural differences between individual infected 3D cOrgs and mock 3D cOrgs, such as differences in the density of cells. Given that there are more large-scale single-cell non-spatial (2D) RNA sequence datasets that had been generated from human post-mortem brain samples, compared to single-cell spatial (3D) RNA sequence datasets from human post-mortem brain samples, the development of data integration approaches by using artificial neural networks such as MLPs, across 2D and 3D single-cell transcriptomics datasets generated from human post-mortem brain samples and human in-vitro systems such as brain organoids is likely to be critical to gain novel insights into neurodegenerative diseases such as AD.