TL;DR
This paper introduces ProGAdNet, a novel algorithm that efficiently infers and compares gene networks under different conditions by jointly analyzing them. This approach significantly improves accuracy and speed compared to existing methods, enabling the identification of key genes and pathways involved in disease states like cancer.
What Problem Does This Paper Address?
Living cells are dynamic systems where gene networks, the intricate web of interactions between genes, can change in response to various factors like environmental shifts, tissue types, or disease progression. Understanding these changes is crucial for deciphering the molecular basis of biological processes and diseases. Current methods often infer gene networks for each condition separately and then compare them. However, this approach overlooks the inherent similarity between networks under different conditions, leading to suboptimal accuracy. The authors propose a more efficient and accurate solution by jointly inferring two gene networks and their differences simultaneously.
Methods
The core of this work is the development of the ProGAdNet (Proximal Gradient Algorithm for Differential Network) algorithm. It tackles the challenge of inferring gene networks from gene expression data, which is often characterized by a high number of potential interactions (genes) and a limited number of samples. ProGAdNet leverages two key properties of gene networks:
- Sparsity: Gene networks are typically sparse, meaning each gene interacts with only a small subset of other genes.
- Similarity: Networks under different conditions are often similar, with only a fraction of interactions changing.
ProGAdNet formulates this problem as an optimization task that jointly infers two gene networks (B and B̃) and their differences. It uses a proximal gradient method to efficiently solve this optimization problem. This method iteratively updates the network parameters, taking advantage of the differentiable and non-differentiable parts of the objective function. A key innovation is the derivation of closed-form solutions for sub-problems within the iterative process, significantly speeding up computation. The algorithm also incorporates stability selection and cross-validation techniques to robustly identify significant network changes and select optimal regularization parameters.
Datasets
The paper evaluates ProGAdNet using two types of datasets:
- Simulated Data:
- Linear Regression Models: Data was generated based on linear regression models to simulate gene expression under two conditions. This allowed for controlled experiments to assess performance metrics like detection power and false discovery rate (FDR) under varying sample sizes and noise levels.
- Gene Networks: The GeneNetWeaver software was used to generate realistic gene network structures. Gene expression data was then simulated based on these networks, mimicking scenarios with different levels of measurement noise and sample sizes. This benchmark is similar to those used in the DREAM5 challenge.
- Real Biological Data:
- TCGA Breast Cancer Data: RNA-seq data from 113 breast tumors and 113 matched normal tissues from The Cancer Genome Atlas (TCGA) database was analyzed.
- TCGA Kidney Cancer Data: RNA-seq data from 72 kidney tumors and 72 matched normal tissues from the TCGA database was also analyzed.
Key Insights and Ideas
- Joint Inference is Superior: Inferring two gene networks simultaneously, rather than separately, significantly improves the accuracy of identifying differential network structures. This is because it exploits the inherent similarity between networks under different conditions.
- Sparsity and Similarity are Key: The ProGAdNet algorithm effectively leverages the sparsity of gene networks and the similarity between networks under different conditions to achieve its performance gains.
- Differential Expression vs. Differential Interaction: The study highlights that genes involved in network changes (differentially interacting genes) are not always the same as differentially expressed genes. This distinction offers a new perspective on understanding disease mechanisms.
- Cancer Signatures: The identified sets of differentially interacting genes (dNet sets) from cancer data show significant enrichment for gene sets related to breast cancer and kidney cancer, suggesting their relevance to disease status.
Key Results
- Improved Accuracy: ProGAdNet demonstrated superior inference accuracy compared to existing methods like Lasso (glmnet) and GENIE3 in simulations, particularly in detecting changed edges.
- Computational Efficiency: ProGAdNet was significantly faster than the lqa algorithm for joint network inference, and also faster than Lasso in some simulation setups. For instance, inferring a linear model with 1000 variables took ProGAdNet 5.82 seconds, compared to 145.15 seconds for lqa.
- Breast Cancer Analysis: Analysis of TCGA breast cancer data identified 268 genes involved in changed network edges. Gene set enrichment analysis revealed that 121 of these genes were significantly enriched in pathways related to breast cancer.
- Kidney Cancer Analysis: Analysis of TCGA kidney cancer data identified 1091 genes involved in changed network edges. The top 460 genes were further analyzed, with many of them being implicated in kidney cancer in existing literature.
Limitations and Future Work
- Parameter Selection: While stability selection and cross-validation are employed, determining the optimal values for regularization parameters (λ₁ and λ₂) can still be computationally intensive.
- Scalability: While ProGAdNet is efficient, the computational burden for very large networks and extensive stability selection might still be a consideration.
- Future Work: The authors suggest that ProGAdNet can be used for further differential network analysis (DiNA) and that the identified gene sets are highly informative for understanding disease status, implying potential for further biological validation and discovery.
Summary updated