Title: De novo inference of cell interactions from single-cell gene expression
Problem or question being addressed
Single-cell spatial transcriptome technology can obtain gene expression and spatial neighboring relationships of cells [1,2]. The spatial adjacency network of cells constitutes a large set of cell interactions, from which we can devise methods to learn the molecular mechanisms leading to interactions and to infer cell interactions. Different from the commonly used methods that rely only on ligand-receptor to infer interactions [3,4,5], we wondered whether this idea could be more general. If there is an AI model whose input is the gene expression of cells, can it predict cell interactions without relying on any priors and assumptions?
Rationale and details of suggested approach
In my previous work, I developed a tool called DeepLinc for denovo reconstruction of cell interaction landscapes from single cell spatial transcriptome [6]. Further, I am currently developing another methodological framework for denovo conference of cell interactions using only single cell gene expression. I borrowed ideas from growing graph modeling (an advanced AI strategies). In short, I regard the neighboring interaction graph of cells as a network generated by a dynamically growing sequential process. When a new cell is added, it interacts with the old cells in the original graph, which is what the model predict (Fig. 1). There is an assumption here that the molecular mechanisms for generating cell interactions are relatively constant. As cell interactions are both cause and effect, whenever a cell with the particular distribution pattern of molecular signatures appears, no matter its order, it will establish a relationship with cells it was supposed to interact with.
Fig. 1 Framework of inferring cell interactions from gene expression without any prior knowledge.
Tests on simulated and real data show that our method can effectively infer cell interactions (AUC~83%), and can predict well across FOVs (field of views) and tissue slices, as well as infer interactions that are not limited to spatial constraints.
How it will affect the broader field
For a long time, inference of cell interaction has been heavily dependent on known cell interaction references (e.g., the expression of ligand-receptors and downstream genes). However, this method will bring unpredictable errors, and has high requirements for the accuracy of ligand annotation. Benchmark studies have shown that many methods of this type infer a large number of false positive interactions, and even the conclusion is completely opposite in severe cases [7]. Furthermore, expression information of ligand-receptors is largely missing in the single-cell spatial transcriptome data. In this context, our method does not rely on any priori for de novo inference of cell interactions based on the latest AI strategies, which is greatly conducive to mining more potential molecular mechanisms. On the other hand, this method can predict cross tissue slices, and infer interactions that are not limited by space.
References
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 Biotechnol 39, 43-58, doi:10.1016/j.tibtech.2020.05.006 (2021).
2 Rodriques, S. G. et al. Slide-seq: A scalable technology for measuring genome-wide expression at high spatial resolution. Science 363, 1463-1467 (2019).
3 Kumar, M. P. et al. Analysis of Single-Cell RNA-Seq Identifies Cell-Cell Communication Associated with Tumor Characteristics. Cell Rep 25, 1458-1468 e1454, doi:10.1016/j.celrep.2018.10.047 (2018).
4 Efremova, M., Vento-Tormo, M., Teichmann, S. A. & Vento-Tormo, R. CellPhoneDB: inferring cell–cell communication from combined expression of multi-subunit ligand–receptor complexes. Nat. Protoc.15, doi:10.1038/s41596-020-0292-x (2020).
5 Browaeys, R., Saelens, W. & Saeys, Y. NicheNet: modeling intercellular communication by linking ligands to target genes. Nat Methods 17, 159-162, doi:10.1038/s41592-019-0667-5 (2020).
6 Li, R. & Yang, X. De novo reconstruction of cell interaction landscapes from single-cell spatial transcriptome data with DeepLinc. Genome Biology 23, 124, doi:10.1186/s13059-022-02692-0 (2022).
7 Dimitrov, D. et al. Comparison of methods and resources for cell-cell communication inference from single-cell RNA-Seq data. Nature Communications 13, 3224, doi:10.1038/s41467-022-30755-0 (2022).