ИСТИНА

Background/Objectives: Commonly used methods for scRNA-seq data analysis allows for precise clustering and cell type identification of blood and nerve tissue cells. However, for connective tissues standard data processing appears insufficient. Application of machine learning and neural networks ensures comprehensive data analysis for cells from subcutaneous adipose tissue. Methods: In addition to canonical set of analysis pipelines including Cell Ranger and the R-package Seurat, we applied machine learning and neural networks. For clustering, we used R-package based on scDeepCluster neural networks, which allows the iterative feedback without adaptive adjustments to feature space. For cell typing we utilized ScCapsNet (CapsNet) neural network. The scCapsNet model extracts specific features (gene expression levels), therefore providing new cell type identification. For precise cell typing, the scCapsNet algorithm mixes the RNA expression levels of two different cell types and then uses the scCapsNet model trained with non-mixed data to predict specific cell types in the mixed data. Results: scDeepCluster application for analysis of MSCs allowed us to identify 8 clusters instead of 6 clusters detected with graph-based clustering algorithm followed by Louvain Modularity Optimization in Cell Ranger count. Using scCapsNet, we revealed additional small clusters with distinctive gene expression profiles and related biochemical processes. Conclusion: Application of machine learning and neural networks ensure complete disclosure of biological insights in contrast to limited application of standard scRNA-seq data analysis algorithms. Grant Reference: Russian Foundation for Basic Research (RFBR) 20-015-00402.