SFINN: inferring gene regulatory network from single-cell and spatial transcriptomic data with shared factor neighborhood and integrated neural network

Abstract

Motivation: The rise of single-cell RNA sequencing (scRNA-seq) technology presents new opportunities for constructing detailed cell type-specific gene regulatory networks (GRNs) to study cell heterogeneity. However, challenges caused by noises, technical errors, and dropout phenomena in scRNA-seq data pose significant obstacles to GRN inference, making the design of accurate GRN inference algorithms still essential. The recent growth of both single-cell and spatial transcriptomic sequencing data enables the development of supervised deep learning methods to infer GRNs on these diverse single-cell datasets.

Results: In this study, we introduce a novel deep learning framework based on shared factor neighborhood and integrated neural network (SFINN) for inferring potential interactions and causalities between transcription factors and target genes from single-cell and spatial transcriptomic data. SFINN utilizes shared factor neighborhood to construct cellular neighborhood network based on gene expression data and additionally integrates cellular network generated from spatial location information. Subsequently, the cell adjacency matrix and gene pair expression are fed into an integrated neural network framework consisting of a graph convolutional neural network and a fully-connected neural network to determine whether the genes interact. Performance evaluation in the tasks of gene interaction and causality prediction against the existing GRN reconstruction algorithms demonstrates the usability and competitiveness of SFINN across different kinds of data. SFINN can be applied to infer GRNs from conventional single-cell sequencing data and spatial transcriptomic data.

Availability and implementation: SFINN can be accessed at GitHub: https://github.com/JGuan-lab/SFINN.

Figure 1.

Overview of SFINN. (A) SFINN utilizes gene pair expression and cell–cell neighborhood graph as inputs for predicting if there is a regulatory relationship between the genes. (B) SFINN fuses the cell–cell adjacency matrix generated by shared factor neighborhood strategy and that generated using cell spatial location. (C) Training and testing data partitioning strategy. The existing ground-truth data are divided into three parts through cross-validation based on the number of TFs, ensuring a 1:1 ratio of positive to negative pairs. Additionally, genes in the testing dataset are strictly separated from genes in the training dataset.

Figure 2.

Performance comparison of SFINN with the existing methods in the task of gene interaction prediction across eight single-cell transcriptomic datasets. (A) The performance of SFINN and seven other GRN methods on eight datasets, each dot denoting the AUROC/AUPRC score for the gene pairs of each transcription factor. (B) The AUPRC/AUROC median across all transcription factors of each method on each dataset.

Figure 3.

Performance comparison of SFINN with the existing algorithms in the task of gene interaction prediction across five spatial transcriptomic datasets. (A) The AUPRC/AUROC median across gene pairs of all transcription factors on each dataset for SFINN and the other five GRN methods. (B) The AUPRC/AUROC median on each dataset of each GRN method.

Figure 4.

Performance comparison of SFINN with the other algorithms applicable for gene causality prediction. The AUPRC/AUROC median across gene pairs of all transcription factors on each dataset for SFINN and the other GRN methods inferring gene causality from (A) single-cell transcriptomic data and (B) spatial transcriptomic data.

Figure 5.

The results of ablation experiments of SFINN. The ablation experiment regarding the learned embedding θ2 from NN in the integrated neural network framework on all datasets for the tasks of (A) gene interaction and (B) causality prediction. The ablation experiment regarding the integration of spatial information (denoted as SI) and expression information (denoted as EI) on all spatial transcriptomic datasets for the tasks of (C) gene interaction and (D) causality prediction.