Novel computational model of gastrula morphogenesis to identify spatial discriminator genes by self-organizing map (SOM) clustering

Abstract

Deciphering the key mechanisms of morphogenesis during embryonic development is crucial to understanding the guiding principles of the body plan and promote applications in biomedical research fields. Although several computational tissue reconstruction methods using cellular gene expression data have been proposed, those methods are insufficient with regard to arranging cells in their correct positions in tissues or organs unless spatial information is explicitly provided. Here, we report SPRESSO, a new in silico three-dimensional (3D) tissue reconstruction method using stochastic self-organizing map (stochastic-SOM) clustering, to estimate the spatial domains of cells in tissues or organs from only their gene expression profiles. With only five gene sets defined by Gene Ontology (GO), we successfully demonstrated the reconstruction of a four-domain structure of mid-gastrula mouse embryo (E7.0) with high reproducibility (success rate = 99%). Interestingly, the five GOs contain 20 genes, most of which are related to differentiation and morphogenesis, such as activin A receptor and Wnt family member genes. Further analysis indicated that Id2 is the most influential gene contributing to the reconstruction. SPRESSO may provide novel and better insights on the mechanisms of 3D structure formation of living tissues via informative genes playing a role as spatial discriminators.

eSPRESSO: a spatial self-organizing-map clustering method for single-cell transcriptomes of various tissue structures using graph-based networks

“…Inspired by Kohonen’s self-organizing-map learning theory 19 , here we extended our previously developed method, SPRESSO 15 , to graph-based networks, which is theoretically applicable to cell-to-cell relationships of any types of topological structures of tissues or organs. The basic algorithm of spatial clustering of cells is a combinatorial optimization to find best gene sets to reproduce known topological network structures of learning objects, or gene vectors of cells.…”

“…We also used the Markov-Chain-Monte-Carlo (MCMC) to optimize the best gene sets to give a maximum reproducibility of a topology of structure. We have also tuned the SOM learning process by introducing stochastically learning (i.e., stochastic-SOM 15 ) that allows learning efficiency in later phase where the extent of learning ability usually decreases monotonously. We refer to this method as eSPRESSO (enhanced-SPRESSO) in this paper.…”

“…Previously, we reported a novel 3D reconstruction method using stochastic-SOM clustering, or SPRESSO (SPatial REconstruction by Stochastic-SOM) 15 , which features gene selections based on Gene Ontology (GO). 16,17 The method yielded high success rates of 3D reconstructions of mouse gastrula embryo and demonstrated a remarkable ability to find SDGs that contribute to differentiation and tissue morphogenesis.…”

Human Cell Atlas and cell-type authentication for regenerative medicine

“…scRNA-seq-based maps have revealed cell-type-specific functions in the liver 26 , blastocyst 27 , and growth plate 28 . The expression of cell adhesion genes and specific gene functions defined by Gene Ontology (GO) terms is being used to develop new tools for single-cell 3D transcriptome analysis that enhance spatial prediction 29 , 30 . Although the identification of cell types in a spatial context is expected to yield more information relevant to the in vivo environment, these cutting-edge approaches are still at the elementary stage and need further improvement before they can be widely used.…”

Hybrid quantum-classical unsupervised data clustering based on the Self-Organizing Feature Map

“…In particular, the self-organizing feature map (SOFM) [12][13][14], is a type of artificial neural network (ANN) that is trained in an unsupervised manner. SOFMs are used in many areas [15][16][17][18][19][20][21][22] and in comparison with many other * Email address:pyrkov@icp.ac.ru artificial neural networks, they apply competitive learning and preserve the topological properties of the input space [23]. The SOFMs represent data in a fundamentally topological way that allows one to perform dimensionality reduction.…”