Recovering gene regulatory networks in single-cell multiomics data with PRISM-GRN

Abstract

Understanding Gene Regulatory Networks (GRNs) is crucial for deciphering cellular heterogeneity and the mechanisms underlying development and disease. However, current GRN inference methods fail to utilize multiomics data and prior knowledge from a biologically-interpretable insight. Therefore, we propose PRISM-GRN, a Bayesian model that seamlessly incorporates known GRNs, along with scRNA-seq and scATAC-seq data, into a probabilistic framework to reconstruct cell type-specific GRNs. PRISM-GRN employs a biologically interpretable architecture firmly rooted in the established gene regulatory mechanism, which asserts that gene expression is influenced by TF expression levels and gene chromatin accessibility through GRNs. Accordingly, PRISM-GRN decomposes observable data into biologically meaningful latent variables through a mechanism-informed generation process and a prior-GRN-primed inference process, enabling precise and robust GRN reconstruction. We evaluate PRISM-GRN on four benchmarking datasets with paired scRNA-seq and scATAC-seq data, demonstrating its superior performance over seven baseline methods in GRN reconstruction, especially its higher precision under the inherently imbalanced scenario where the true regulatory interaction is sparse. Furthermore, benchmarking on directed GRNs highlights PRISM-GRN's ability to capture causality in gene regulation derived from the biologically-interpretable architecture. More importantly, PRISM-GRN performs well with unpaired omics data and limited prior GRN information, showcasing its flexibility and adaptability across various biological contexts. Finally, biological analyses on PBMC datasets demonstrate PRISM-GRN's potential to facilitate the identification of cell type-specific or context-specific GRNs across broader real-world biological research applications. Overall, PRISM-GRN provides a novel paradigm for precise, robust, and interpretable exploration of causal GRNs with prior knowledge and multiomics data.