TL;DR
This paper introduces SiCNet, a novel method for constructing causal gene regulatory networks (GRNs) at the individual cell level. By leveraging causal inference and a reference network, SiCNet generates cell-specific networks that improve cell clustering and reveal dynamic biological processes like cancer progression and cellular reprogramming.
What Problem Does This Paper Address?
Understanding the complex regulatory mechanisms that drive cellular behavior, development, and disease is a major challenge in biology. Gene Regulatory Networks (GRNs) are key to deciphering these mechanisms, but existing methods often struggle with inferring causality, dynamic rewiring, and context specificity, especially at the single-cell level. Current approaches typically infer GRNs at the cell-type level, failing to capture the heterogeneity and unique regulatory logic of individual cells. This limits our ability to understand intricate biological processes that unfold dynamically within a population of cells.
Methods
SiCNet employs a two-step approach:
- Reference Causal Network Construction (RCN): An initial GRN is built using a reference scRNA-seq dataset. This involves pairwise causal inference between genes using a regression-based method and 5-fold cross-validation. A Causal Strength Index (CSI) is calculated to quantify the causal effect between gene pairs. A protein-protein interaction (PPI) network from STRING is integrated as prior knowledge.
- Cell-Specific Causal Network Inference (CCN): For each individual cell, SiCNet uses its gene expression profile and the previously constructed RCN to infer cell-specific causal relationships. This involves predicting gene expression with and without potential regulators and assessing the change in prediction error.
- Network Outdegree Matrix (ODM): The cell-specific causal networks are then transformed into an ODM. This matrix represents the higher-order regulatory information of each gene (outdegree) and can be used as input for standard scRNA-seq analysis tools like clustering and dimensionality reduction, replacing the original gene expression matrix (GEM).
Datasets
The authors evaluated SiCNet using several publicly available scRNA-seq datasets:
- For benchmarking clustering performance:
- Biase dataset (mouse embryos)
- Enge dataset (human pancreas)
- Darmanis dataset (human brain)
- Fink dataset (human stromal cells)
- Goolam dataset (mouse embryo cells)
- For demonstrating biological applications:
- Human colorectal cancer scRNA-seq dataset (healthy adjacent, border, and core regions)
- Mouse embryonic fibroblast (MEF) to induced pluripotent stem cell (iPSC) reprogramming dataset (6 time points)
- Human bone marrow mononuclear cell (BMMC) dataset (hematopoiesis)
Key Insights and Ideas
- Capturing Individual Cell Logic: SiCNet moves beyond cell-type level GRNs to reveal the unique regulatory logic of each individual cell, crucial for understanding cellular heterogeneity.
- Causality Over Correlation: The method explicitly focuses on inferring causal relationships, providing a more robust understanding of gene interactions than correlation-based methods.
- ODM for Enhanced Analysis: Transforming cell-specific networks into an ODM significantly improves the clarity and separation of cell populations in dimensionality reduction and clustering analyses.
- Dynamic Process Unveiling: SiCNet effectively models dynamic biological processes like cancer progression and cellular reprogramming by capturing transient states and regulatory shifts.
- Identifying Key Regulators: The method can pinpoint critical regulators driving these complex biological transitions.
Key Results
- Improved Clustering: Visualizations (t-SNE) of ODMs showed clearer demarcations between cell types compared to GEMs across multiple datasets.
- Quantitative Clustering Performance: ODMs consistently outperformed GEMs in clustering performance metrics (ARI, AMI, NMI, FMI) across various algorithms (SC3, CIDR, Scanpy, SIMLR, Supercell) on benchmark datasets. For example, in the Enge dataset, ODM achieved higher ARI and AMI scores with SC3 compared to GEM.
- Colorectal Cancer Progression: SiCNet identified 26 regulators with enhanced activity from healthy to core tumor regions, many of which are linked to known oncogenic pathways (e.g., ABL1, NCOA1, ATF3, JUN, EGR2). The method also identified critical states in the transitional border region.
- Cellular Reprogramming Dynamics: SiCNet delineated distinct cell-type specific networks during MEF to iPSC reprogramming, identifying key regulators for different stages and uncovering dynamic regulatory processes like cell cycle progression and DNA repair.
- Hematopoiesis Trajectories: The method successfully identified regulatory dynamics along B cell, monocyte, and erythroid cell lineages, highlighting regulators like CD74, IRF3, JAK1, CEBPB, and KIT.
Limitations and Future Work
- Assumptions: The current SiCNet method assumes steady states and does not explicitly account for nonlinearity, non-steady-state conditions, or mechanistic models.
- Data Modalities: The approach currently focuses on gene expression data and does not integrate other potentially informative data modalities like unspliced mRNAs, DNA methylation, or chromatin accessibility.
- Future Directions: The authors suggest extending SiCNet to spatial transcriptomics data and applying it to sample-specific network construction. They also highlight the potential for uncovering novel regulatory pathways and identifying targets for perturbation.
Summary updated