iMIRACLE: an Iterative Multi-View Graph Neural Network to Model Intercellular Gene Regulation from Spatial Transcriptomic Data

Abstract

Spatial transcriptomics has transformed genomic research by measuring spatially resolved gene expressions, allowing us to investigate how cells adapt to their microenvironment via modulating their expressed genes. This essential process usually starts from cell-cell communication (CCC) via ligand-receptor (LR) interaction, leading to regulatory changes within the receiver cell. However, few methods were developed to connect them to provide biological insights into intercellular regulation. To fill this gap, we propose iMiracle, an iterative multi-view graph neural network that models each cell's intercellular regulation with three key features. Firstly, iMiracle integrates inter- and intra-cellular networks to jointly estimate cell-type- and micro-environment-driven gene expressions. Optionally, it allows prior knowledge of intra-cellular networks as pre-structured masks to maintain biological relevance. Secondly, iMiracle employs iterative learning to overcome the sparsity of spatial transcriptomic data and gradually fill in the missing edges in the CCC network. Thirdly, iMiracle infers a cell-specific ligand-gene regulatory score based on the contributions of different LR pairs to interpret inter-cellular regulation. We applied iMiracle to nine simulated and eight real datasets from three sequencing platforms and demonstrated that iMiracle consistently outperformed ten methods in gene expression imputation and four methods in regulatory score inference. Lastly, we developed iMiracle as an open-source software and anticipate that it can be a powerful tool in decoding the complexities of inter-cellular transcriptional regulation.

Keywords: cell–ell communications; graph neural networks; inter-cellular gene regulation; spatial transcriptomics.

Figure 7:

Additional regulatory score inference results.

Figure 1:

Overview of iMiracle. iMiracle initiates with a sparse cell-by-gene matrix ${\mathrm{X}}_{obs}\in {\mathbb{R}}^{n\times m}$ ($n$ cells and $m$ genes), spatial coordinates $\mathbf{C}\in {\mathbb{R}}^{n\times 2}$, and cell type information $\mathbf{T}\in {\mathbb{R}}^{n\times t}$ ($t$ cell types). It constructs multi-view cell-cell communication networks ${\text{G}}_{\text{C}}$ to model various ligand-receptor interactions. Node embeddings for each cell are generated per view through a graph neural network. A multilayer perceptron then decodes gene expression ${\stackrel{\mathbf{ˆ}}{\mathbf{X}}}_s$ influenced by these LR interactions, integrating knowledge from an established gene regulatory network. iMiracle isolates the baseline gene expression matrix ${\stackrel{\mathbf{ˆ}}{\mathbf{X}}}_b$ solely determined by cell types. The final imputed gene expression matrix $\stackrel{\mathbf{ˆ}}{\mathbf{X}}$ merges the baseline matrix with expressions from ligand-receptor interactions. Through iterative learning, $\stackrel{\mathit{ˆ}}{\mathbf{X}}$ is used to progressively refine the multi-view graph, enhancing both imputation precision and the inference of ligand-to-gene regulatory scores.

Figure 2:

iMiracle fully delineates the CCC network via iterative learning, uncovering up to 181% more LR interactions in 10x Visium, 67% in Stereoseq, and 153% in SlideseqV2.

Figure 3:

iMiracle consistently outperforms other methods in the regulatory score inference. (A) In this simulation, 1000 cells are spatially arranged in a 100-unit square. Cell types were determined by their locations, incorporating high-expression zones for various LR pairs, to realistically model gene expression and CCC dynamics. This setup is utilized for inferring cellular-level regulatory scores. (B) Benchmarking results demonstrate iMiracle’s superior accuracy in inferring ligand-gene regulatory scores, surpassing all four baselines across all four metrics.

Figure 4:

iMiracle reveals substantial regulatory heterogeneity across cells. (A) Detailed ground truth segmentation of the cortical layers and white matter (WM) within the DLPFC section of sample 151507. (B) Visualization of the observed expression pattern of GJA1. (C) Prediction of the baseline expression profile for GJA1. (D) Prediction of the LR-regulated expression for GJA1. (E) Identification of three key regions within sample 151507. (F-H) Top three LR interactions and their corresponding regulatory scores. LR pairs PTN-SDC4 and LRRC4B-PTPRD were discovered through an iterative learning approach, indicated by their red colors. (I) Heatmap illustration of the percentage of cells featuring the top five LR pairs in each identified region. (J-L) Jaccard similarity of the top five LR pairs for cells within each region, revealing substantial regulatory heterogeneity across cells.

Figure 5:

Variant analysis. (A) RMSE w.r.t. different variants of iMiracle for gene imputation. (B) Four regulatory score inference metrics w.r.t. variants of iMiracle. (C) Four regulatory score inference metrics w.r.t. GNN architectures.

Figure 6:

Parameter analysis. (A) Precision w.r.t. different blending coefficient $\alpha$. (B) Precision w.r.t. different hidden dimension $d$. (B) Precision w.r.t. different number of neighbors $k$ for graph construction. (B) Precision w.r.t. GNN layers $L$.