MulNet: a scalable framework for reconstructing intra- and intercellular signaling networks from bulk and single-cell RNA-seq data

Mingfei Han, Xiaoqing Chen, Xiao Li, Jie Ma, Tao Chen, Chunyuan Yang, Juan Wang, Yingxing Li, Wenting Guo, Yunping ZhuBriefings in Bioinformatics 2025 · DOI

TL;DR

This paper introduces MulNet, a novel framework that reconstructs complex intra- and intercellular signaling networks by integrating diverse molecular interactions with gene expression data. MulNet uses a two-level reinforcement learning approach to identify gene modules and prognostic regulators, demonstrating superior performance over existing methods in identifying biologically relevant networks from both bulk and single-cell RNA-seq data.

What Problem Does This Paper Address?

Understanding how genes are regulated and how cells communicate is fundamental to deciphering complex biological processes, especially in diseases like cancer. Current methods often focus on specific types of molecular interactions (e.g., gene regulatory networks or protein-protein interactions) in isolation, leading to a fragmented view of cellular signaling. This fragmentation limits our ability to fully grasp the intricate interplay of molecules that drive cellular behavior and disease progression. Specifically, inferring comprehensive networks that capture both within-cell (intra-cellular) and between-cell (inter-cellular) communication from the increasingly abundant but complex gene expression data (both bulk and single-cell) remains a significant challenge.

Methods

MulNet tackles this challenge with a three-step framework:

  1. Multilayer Network Construction: MulNet integrates gene expression data with various reference interaction datasets (protein-protein interactions, TF-target interactions, miRNA-target interactions). It constructs a "multilayer network" where different types of interactions form distinct layers. To enhance robustness, interactions supported by at least two independent databases are prioritized.
  2. Gene Module Identification: This is the core innovation. MulNet employs a two-level reinforcement learning (RL) approach.
  3. Low-level controller: Assigns functional units (individual genes or combinations of interacting genes) to modules based on correlation probabilities, aiming to maximize cumulative correlation.
  4. High-level controller: Optimizes module structure by deciding whether to split existing modules or remove less reliable ones, using the low-level controller's output as a reward signal. This hierarchical RL approach allows for dynamic module discovery and refinement.
  5. Network-Based Survival Analysis: MulNet integrates patient survival data to identify prognostic regulators and their downstream targets. This method leverages the co-expression patterns within identified modules, offering a more sensitive approach than traditional expression-based survival analysis, especially for low-expression regulators like miRNAs.

Datasets

MulNet was evaluated using:

  • TCGA pan-cancer data: Bulk RNA-seq data from multiple cancer types for benchmarking against existing methods.
  • Human scRNA-seq datasets: Data from 10 cancer types (breast, head and neck, prostate, etc.) for evaluating performance on single-cell data.
  • Colon adenocarcinoma (COAD) RNA-seq data: Used for a case study to reconstruct intra-cellular signaling networks and identify prognostic regulators.
  • Head and neck squamous cell carcinoma (HNSCC) scRNA-seq data: Used for a case study to reconstruct inter-cellular communication networks between fibroblasts and malignant cells.

The paper also leverages publicly available reference interaction databases (STRING, BIOGRID, DIP, HURI, MINT, hTFtarget, RegNetwork, TFLink, TRRUST, miRDB, miRTarBase) and known cancer modules from previous studies for benchmarking.

Key Insights and Ideas

  • The power of integration: By combining diverse interaction types into a multilayer network, MulNet captures a more holistic view of gene regulation than methods focusing on single interaction types.
  • Hierarchical RL for network discovery: The two-level RL framework is a clever way to balance the exploration of new module structures with the refinement of existing ones, leading to more robust gene module identification.
  • Beyond expression: Network context matters for survival: Network-based survival analysis, which considers how genes interact within modules, is significantly more powerful than analyzing individual gene expression levels for identifying prognostic biomarkers, especially for challenging targets like miRNAs.
  • Uncovering novel therapeutic targets: The framework successfully identified known cancer regulators and, importantly, proposed novel candidates like miR-8485, which were then experimentally validated to inhibit tumor growth.

Key Results

  • Superior module detection: MulNet consistently outperformed state-of-the-art methods in identifying biologically relevant gene modules on both bulk and single-cell RNA-seq data across multiple cancer types.
  • Identification of novel regulators: In colon cancer, MulNet identified miR-8485 as a promising therapeutic target, with experimental validation showing its ability to inhibit proliferation, migration, and tumor growth.
  • Reconstruction of intercellular communication: In head and neck cancer, MulNet elucidated complex signaling networks between cancer-associated fibroblasts and malignant cells, revealing key TFs and ligands involved in tumor progression and metastasis.
  • High sensitivity for prognostic markers: Network-based survival analysis identified a significantly higher number of prognostic regulators (TFs and miRNAs) compared to conventional expression-based methods, with a large proportion of these targets being missed by traditional approaches.

Limitations and Future Work

  • Species specificity: MulNet's reliance on extensive reference interaction data makes it most effective for species with well-annotated genomes (like humans and mice). Its applicability to species with less available data (e.g., E. coli, yeast) is limited, though this is expected to improve as more interaction data becomes available.
  • Scalability: While described as scalable, the computational demands of multilayer network construction and RL-based analysis for very large datasets could be a consideration.
  • Future directions: The authors suggest that the framework's flexibility in integrating user-defined interactions could lead to broad applications in various omics studies. Further preclinical and clinical studies are warranted to validate findings like the therapeutic potential of miR-8485.

Summary updated

BibTeX
@article{han2025mulnet,
  author = {Mingfei Han and Xiaoqing Chen and Xiao Li and Jie Ma and Tao Chen and Chunyuan Yang and Juan Wang and Yingxing Li and Wenting Guo and Yunping Zhu},
  title = {{MulNet: a scalable framework for reconstructing intra- and intercellular signaling networks from bulk and single-cell RNA-seq data}},
  journal = {Briefings in Bioinformatics},
  year = {2025},
  doi = {10.1093/bib/bbaf081},
}
Figure