TL;DR
This paper introduces AJGM, a novel method for inferring gene networks from heterogeneous biological data, like single-cell RNA sequencing (scRNA-seq). AJGM simultaneously classifies samples into subtypes and learns their distinct gene networks, while also identifying shared relationships across all subtypes. This approach improves accuracy by adaptively adjusting network similarities and incorporating an "overall" network that captures commonalities.
What Problem Does This Paper Address?
Understanding how genes interact within biological systems is crucial for deciphering cellular functions and disease mechanisms. Gene networks, which represent these interactions, are often studied using gene expression data. However, biological samples, especially from single-cell experiments, can exhibit significant heterogeneity. This means different groups of cells (subtypes) might have distinct gene expression patterns and, importantly, different underlying gene interaction networks.
Traditional methods for inferring gene networks often struggle with this heterogeneity. Some assume all samples belong to a single network, ignoring subtype-specific differences. Others try to classify samples first and then infer networks, but this can lead to suboptimal results if the classification is inaccurate or if the networks are not well-aligned. A key challenge is that both shared (common across subtypes) and subtype-specific relationships are biologically important. Furthermore, scRNA-seq data is notoriously "zero-inflated" (many genes have zero expression counts), which violates the assumptions of many standard statistical models, including Gaussian Graphical Models (GGMs).
This paper tackles these issues by proposing a method that can:
- Simultaneously classify samples into subtypes and infer their respective gene networks.
- Adaptively adjust the similarity between these subtype networks, acknowledging that not all relationships will be equally similar across subtypes.
- Leverage an "overall" gene network to better capture relationships shared by all samples, thereby improving the accuracy of both shared and subtype-specific network inference.
- Handle the zero-inflation problem inherent in scRNA-seq data.
Methods
The core of AJGM is an Adaptive Joint Graphical Model (AJGM) that builds upon Gaussian Graphical Models (GGMs). Here's a breakdown of its key components:
- Mixture Gaussian Model for Subtypes: AJGM models the gene expression data as coming from a mixture of $K$ Gaussian distributions, where each component represents a subtype. This allows for simultaneous classification of samples and estimation of subtype-specific precision matrices ($\Omega_k$), which represent the gene networks.
- Overall Gene Network: To capture shared relationships, AJGM introduces an "overall" gene network ($\Omega_\chi$). This network is learned from all samples, irrespective of their subtype, and is designed to identify connections that are consistent across the majority of samples. This overall network then influences the estimation of subtype networks.
- Joint Estimation with Adaptive Weights: The model jointly estimates all subtype networks and the overall network. Crucially, it uses a fused penalty term that incorporates adaptive weights. These weights dynamically adjust the degree of similarity enforced between subtype networks and between subtype networks and the overall network. This allows the model to focus more on shared relationships when they are strong and less when they are weak or inconsistent.
- Data Imputation for Zero-Inflation: To address the zero-inflation in scRNA-seq data, AJGM incorporates a data imputation step within its Expectation-Maximization (EM) algorithm. This imputation is performed at each iteration, using subtype-specific Gaussian distributions to fill in missing (zero) values, making the data more amenable to GGM assumptions.
- Expectation-Maximization (EM) Algorithm and ADMM: The parameter estimation is performed using an EM algorithm. For updating the precision matrices, AJGM employs the Alternating Direction Method of Multipliers (ADMM) algorithm, a common and efficient technique for solving such optimization problems, especially in high-dimensional settings.
What's novel here is the combination of:
- Adaptive similarity constraints that go beyond simple group lasso penalties.
- The explicit inclusion of an "overall" network to guide subtype network inference.
- The integrated data imputation within the EM framework for zero-inflated data.
Datasets
The paper evaluates AJGM on both simulated and real-world biological datasets:
- Simulated Data: Datasets were generated with varying numbers of subtypes ($K=3$ and $K=4$) and dimensions ($p=100$ genes), incorporating different types of network modules (dense, circle, star, and zero modules) and a mechanism for generating "public" (shared) blocks within precision matrices. Zero-inflation was also simulated.
- Real Application 1: Triple-Negative Breast Cancer (TNBC) scRNA-seq Data: Data from 220 TNBC cells (GEO accession number GSE75688) was used. Genes were selected based on the PAM50 gene set and known breast cancer hub genes.
- Real Application 2: TNBC Bulk RNA-seq Data: Data from 108 TNBC samples from The Cancer Genome Atlas (TCGA) was used to demonstrate AJGM's applicability to bulk data.
- Real Application 3: Arabidopsis Root Procambium-Type Cells Time-Series Data: Data from 10 distinct time points (GSE152766) was used to test AJGM's scalability and ability to handle time-series data by reclassifying cells into early, middle, and late stages.
- Real Application 4: Mouse Embryonic Stem Cells Data: Data from two time points (GSE65525) was used to study differentiation processes regulated by leukemia inhibitory factor (LIF).
The paper does not introduce a new benchmark dataset but rather uses existing, well-characterized datasets to validate its method against established approaches.
Key Insights and Ideas
- Heterogeneity is Key: The paper strongly emphasizes that biological systems are rarely uniform. Different cell subtypes can have fundamentally different regulatory mechanisms, and ignoring this leads to incomplete or inaccurate network models.
- Shared vs. Specific is a Powerful Distinction: By explicitly modeling both shared and subtype-specific relationships, AJGM provides a more nuanced understanding. Shared relationships offer insights into core biological functions, while subtype-specific ones highlight unique cellular behaviors and potential therapeutic targets.
- The "Overall" Network as a Guiding Light: The introduction of an overall network is a clever way to provide a global context for subtype network inference. It acts as a prior, nudging the model towards identifying connections that are likely to be present across the board, thus improving the robustness of the entire inference process.
- Adaptive Weights are Crucial: The idea that not all subtype networks should be forced to be equally similar is intuitive. Adaptive weights allow the model to learn the degree of similarity, making it more flexible and accurate than methods with fixed similarity constraints.
- Zero-Inflation is a Major Hurdle for scRNA-seq: The paper highlights that standard GGM assumptions are often violated by scRNA-seq data. Their integrated imputation method is a practical solution to make these powerful models applicable to this challenging data type.
Key Results
- Superior Performance on Simulated Data: AJGM consistently outperformed several state-of-the-art methods (BLGGM, GGMPF, MGGM, SCGGM, SILGGM) in terms of False Positive Rate (FPR), True Positive Rate (TPR), and F1 scores for subtype network inference across various scenarios, including different numbers of subtypes and zero-inflated data.
- Accurate Shared Relationship Inference: AJGM demonstrated a remarkable ability to accurately estimate shared relationships, often outperforming other methods in identifying these crucial connections (as shown in Figure 3).
- Biologically Meaningful Findings in TNBC: When applied to TNBC scRNA-seq data, AJGM identified biologically relevant hub genes (e.g., FOXA1, FGFR4, CDC20, MAP3K1, MYC) and enriched pathways (e.g., mitotic cell cycle, cell cycle checkpoint signaling) that align with known TNBC characteristics and prognosis.
- Scalability and Applicability: AJGM showed computational advantages over comparable methods (MGGM) and demonstrated its applicability to both scRNA-seq and bulk RNA-seq data, as well as time-series data, with results showing high consistency with prior annotations.
- Improved Sample Classification: The joint classification and network inference approach led to accurate sample classification, with ARI scores of 1.00 in many simulation scenarios.
Limitations and Future Work
- Challenges with High Zero-Inflation: The paper acknowledges that while data imputation is effective, it can be challenging for scRNA-seq data with extremely high proportions of zeros. The iterative imputation and re-estimation process can be computationally intensive and potentially unstable.
- Overall Network Definition: The authors note that the current definition of the overall network might include relationships present in a majority of samples but not necessarily all samples. They suggest future work could involve separately estimating common and non-common parts of subtype networks for a more precise overall network.
- Computational Efficiency for Very High Dimensions: While AJGM shows computational advantages, future work aims to enhance its ability to handle even higher-dimensional network estimation.
- Future Directions:
- Gibbs Sampling for Imputation: Exploring Gibbs sampling for imputation might offer more robust convergence properties for highly zero-inflated data.
- Improved Overall Network Estimation: Developing methods to more strictly define the overall network to include only truly universal relationships.
- Higher-Dimensional Networks: Extending AJGM to handle even larger gene networks.
- Integration of Spatial and Temporal Data: The authors are exploring extensions to integrate spatial transcriptomics data and temporal sequencing data for a more comprehensive, multi-dimensional biological analysis.
Summary updated