TL;DR
This paper introduces VMPLN, a novel method for inferring gene regulatory networks (GRNs) from single-cell RNA sequencing (scRNA-seq) data, particularly when cells come from mixed populations with unknown identities. VMPLN jointly performs cell clustering and GRN inference, outperforming existing methods, especially in challenging scenarios with high cell type mixing.
What Problem Does This Paper Address?
Understanding how genes interact within a cell to control its behavior is crucial for deciphering complex biological systems. These interactions form gene regulatory networks (GRNs). Traditionally, GRNs were inferred from bulk RNA sequencing data, which averages gene expression across many cells, losing individual cell-level detail. The advent of single-cell RNA sequencing (scRNA-seq) allows for gene expression measurement at the individual cell level, opening up new possibilities for GRN inference. However, scRNA-seq data often contains cells from multiple, distinct cell types, and their identities are unknown. A common approach is to first cluster cells into types and then infer GRNs for each cluster separately. This two-step process can lead to inaccurate GRN estimates if cell types are not well-separated or if there's significant overlap (high mixing degree). This paper addresses this challenge by proposing a method that simultaneously clusters cells and infers GRNs, thereby handling mixed cell populations more effectively.
Methods
The core of the proposed method is the Mixture Poisson Log-Normal (MPLN) model. This model is designed for count data (like scRNA-seq) and can handle the over-dispersion often observed in such data. The MPLN models gene expression as a Poisson distribution conditional on a latent variable, which in turn follows a normal distribution. The precision matrix of this latent variable's normal distribution represents the GRN for a specific cell type. The MPLN extends this by assuming a mixture of these models, where each component represents a different cell type, and the precision matrices of these components are the GRNs for those cell types.
To overcome the computational intractability of directly optimizing the MPLN's log-likelihood, the authors developed a Variational Mixture Poisson Log-Normal (VMPLN) algorithm. This algorithm uses variational inference to approximate the complex posterior distribution and jointly estimates cell type proportions, cell type-specific GRNs, and cell identities. A key algorithmic innovation is the use of an Alternating Direction Method of Multipliers (ADMM) algorithm to efficiently optimize the parameters related to the GRNs.
Datasets
The paper uses several datasets for evaluation:
- Simulated datasets: Generated based on the proposed MPLN model and a related mixture multinomial log-normal distribution to test performance under various conditions (mixing levels, dropout rates, dimensions, graph structures).
- Real scRNA-seq datasets:
- Kang dataset: Human peripheral blood mononuclear cells (PBMC) from Ref. [28] (7217 cells, 10 cell types).
- Zheng dataset: Human PBMC cells from Ref. [29] (7424 cells, 6 cell types).
- Liao dataset: Macrophage cells from patients with SARS-CoV-2 infection from Ref. [33] (29,980 cells, 8 patients, 4 macrophage clusters).
- SARS-CoV-2 infection data: Single-cell RNA-seq data with SARS-CoV-2 infection information for each cell from Ref. [34], used to validate findings related to UPR and ER stress genes.
Key Insights and Ideas
- Joint inference is crucial: The paper strongly advocates for jointly inferring cell types and GRNs, demonstrating that this approach significantly improves GRN accuracy, especially when cell types are highly mixed.
- MPLN for scRNA-seq: The Poisson log-normal model is well-suited for scRNA-seq data due to its ability to handle count data and over-dispersion, and its extension to a mixture model effectively captures heterogeneity across cell types.
- ADMM for efficient optimization: The development of an ADMM-based optimization strategy for the variational parameters makes the VMPLN algorithm computationally feasible for large scRNA-seq datasets.
- COVID-19 insights: The application of VMPLN to COVID-19 data revealed potential differences in GRNs of alveolar macrophages between moderate and severe infection cases, highlighting the role of NR4A1 and UPR/ER stress pathways.
Key Results
- Superior performance in simulations: VMPLN consistently outperformed state-of-the-art methods (including VPLN, LPGM, Glasso, PPCOR, GENIE3, PIDC, CSN, SINCERITIES) in terms of pAUPRC ratio and early precision ratio across various simulation scenarios, particularly those with high cell mixing degrees.
- Robustness in real data: Benchmarking on two real scRNA-seq datasets (Kang and Zheng) showed VMPLN achieving the highest pAUPRC and early precision ratios in a significant number of cell types, demonstrating its effectiveness on complex biological data.
- COVID-19 GRN analysis: VMPLN identified critical differences in regulatory networks in immune cells between patients with moderate and severe SARS-CoV-2 symptoms, pointing to NR4A1 as a key regulator in severe cases and implicating UPR and ER stress pathways.
- Computational efficiency: While VMPLN incurs substantial time costs, its performance is justified. The ADMM algorithm helps maintain time costs comparable to other methods, and it supports parallel computation. Memory usage was found to be moderate.
Limitations and Future Work
- Linearity assumption: VMPLN assumes linear regulatory relationships. The authors suggest future work on developing methods that can account for nonlinear regulatory relationships, potentially by integrating with methods like GENIE3 or by developing new nonlinear models.
- Incorporating multi-omics data: Currently, VMPLN is designed for scRNA-seq data. Future work could explore integrating other single-cell omics data (e.g., epigenomics, proteomics) to improve sensitivity and reduce false discoveries.
- Determining cell type number: A significant challenge in analyzing scRNA-seq data is determining the optimal number of cell types. The authors suggest using pre-clustering, biological priors, or incorporating Dirichlet process mixtures to jointly infer the cell type number.
Summary updated