Title: De novo reconstruction of cell interaction landscapes from single-cell spatial transcriptome data with DeepLinc
Problem or question being addressed
As the frontier of technological development in recent years, single-cell spatial omics technology provides both the spatial localization of cells in physiological tissues and various quantitative omics data, bringing an unprecedented basis for understanding the assembly and interaction between heterogeneous cells in tissues .
However, considering the complexity of single-cell spatial omics data (abundant data points, low signal-to-noise ratio, high dimensionality and sparsity, multi-factor nonlinear correlation...) , how to design targeted and efficient algorithms to extract representative information from the limited snapshots of cell spatial omics, mine the hidden correlations between intracellular molecular features and intercellular interactions, and systematically reconstruct a more complete cell interaction network, is an important frontier issue in the field of single-cell spatial omics, the theoretical and technical research direction of cell interaction analysis.
In traditional single-cell omics studies, the identification of cellular interactions often relies on known ligand-receptor relationships, but the mechanisms mediated by ligand-receptors are far from covering all interaction patterns . In addition, in the spatial transcriptome data at single-cell resolution, the ligand/receptor expression information of most cells are often incomplete, which does not meet the requirements of identifying cellular interactions . These factors make the application of conventional analysis methods very difficult and limited, calling a de novo algorithmic framework.
Rationale for your approach
The intrinsic gene expression profile of each single cell is both a consequence and a defining factor of the complicated cell interaction network in physiological contexts [4,5]. Artificial intelligence has inherent advantages in resolving such complex bidirectional relationships. For example, deep generative models have been proven to be powerful tools for leveraging latent features and modeling high-dimensional scRNA-seq data [6,7]. In the present study, we adapted another type of deep generative model, variational graph autoencoders (VGAEs) , for encoding cell-cell interaction features from spatial single-cell transcriptome data and eventually regenerating full cell-cell interaction landscapes. Specifically, DeepLinc assumes that the neighboring cells should be much more likely to have some types of interactions than randomly picked non-neighboring cells that are far away from each other. Therefore, the spatial adjacent network of cells becomes the learned object for our model to mine the intrinsic relationship between intracellular molecular features and cell interactions.
Details of suggested approach
Based on a novel artificial intelligence strategy, we have developed a bioinformatics tool DeepLinc, which uses single-cell spatial transcriptome data to reconstruct the cellular interaction network in physiological tissues. Specifically, cell adjacency network (A) and gene expression profiles (X) are fed into the VGAE consisting of two graph convolutional layers. As the output of this variational graph convolutional network (VGCN) encoder, the latent representation (H) captures the characteristics of a single cell itself and its neighboring cells. In addition, H is further constrained by an adversarial regularization module from a prior Gaussian distribution (Fig. 1). Next, with the information learned above, the decoder performs a dot product operation on H to generate a new adjacency matrix (A’) presenting the reconstructed cell-cell interaction network (Fig. 1). On the other hand, the vectors of H, which represent the latent information of cell interaction landscapes and gene expression profiles, could be extracted for the visualization and clustering of single cells.
DeepLinc demonstrates high efficiency in learning from imperfect and incomplete spatial transcriptome data, filtering false interactions, and inferring missing distal and proximal interactions . The reconstructed full networks of cell interactions exhibited high physiological relevance. Interrogations of the DeepLinc pipeline reveal signature genes that are potentially involved in shaping cell interaction landscapes.
Fig. 1 Schematic description of the DeepLinc pipeline. Adapted from Li et al., 2022, Genome Biol.
How it will affect the broader field
In summary, as a tool for deep data mining, DeepLinc further empowers rapidly emerging spatial transcriptome profiles for the de novo reconstruction of cell interaction maps. This new strategy, facilitated by deep graph convolutional networks, does not rely on any prior knowledge of cells or gene functions. It can fully release the advantages and potential of single-cell spatial omics analysis technology by mining the multi-level complex correlations between cellular molecular profiles and complex physiological and pathological tissues. In addition, the framework also considers the scalability of multimodal data, such as morphological data, new omics data, time series information, etc. Therefore, a scalable framework is an important scientific value of this method. Furthermore, the method could promote new biological discoveries by identifying key molecular features closely related to cell interactions and serving as an informative resource for follow-up research. We anticipate that the combination of state-of-the-art spatial transcriptome profiling techniques and an efficient data deep mining framework will greatly facilitate the identification of the biological mechanisms underlying complex cell communication networks with physiological relevance.
1 Liao, J., Lu, X., Shao, X., Zhu, L. & Fan, X. Uncovering an Organ's Molecular Architecture at Single-Cell Resolution by Spatially Resolved Transcriptomics. Trends Biotechnol39, 43-58, doi:10.1016/j.tibtech.2020.05.006 (2021).
2 Burgess, D. J. Spatial transcriptomics coming of age. Nat. Rev. Genet.20, doi:10.1038/s41576-019-0129-z (2019).
3 Armingol, E., Officer, A., Harismendy, O. & Lewis, N. E. Deciphering cell-cell interactions and communication from gene expression. Nat Rev Genet22, 71-88, doi:10.1038/s41576-020-00292-x (2021).
4 Chen, X. & Cubillos-Ruiz, J. R. Endoplasmic reticulum stress signals in the tumour and its microenvironment. Nat Rev Cancer21, 71-88, doi:10.1038/s41568-020-00312-2 (2021).
5 Buckley, C. D., Ospelt, C., Gay, S. & Midwood, K. S. Location, location, location: how the tissue microenvironment affects inflammation in RA. Nat Rev Rheumatol, doi:10.1038/s41584-020-00570-2 (2021).
6 Eraslan, G., Simon, L. M., Mircea, M., Mueller, N. S. & Theis, F. J. Single-cell RNA-seq denoising using a deep count autoencoder. Nat Commun10, 390, doi:10.1038/s41467-018-07931-2 (2019).
7 Xu, Y. et al. scIGANs: single-cell RNA-seq imputation using generative adversarial networks. Nucleic Acids Res48, e85, doi:10.1093/nar/gkaa506 (2020).
8 Kipf, T. N. & Welling, M. Variational Graph Auto-Encoders. (2016).
9 Li, R. & Yang, X. De novo reconstruction of cell interaction landscapes from single-cell spatial transcriptome data with DeepLinc. Genome Biology23, 124, doi:10.1186/s13059-022-02692-0 (2022).