Two-way sparsity for time-varying networks with applications in genomics

Thomas E. Bartlett, Ioannis Kosmidis, Ricardo SilvaThe Annals of Applied Statistics 2021 · DOI · arXiv

Abstract

We propose a novel way of modelling time-varying networks by inducing two-way sparsity on local models of node connectivity. This two-way sparsity separately promotes sparsity across time and sparsity across variables (within time). Separation of these two types of sparsity is achieved through a novel prior structure which draws on ideas from the Bayesian lasso and from copula modelling. We provide an efficient implementation of the proposed model via a Gibbs sampler, and we apply the model to data from neural development. In doing so, we demonstrate that the proposed model is able to identify changes in genomic network structure that match current biological knowledge. Such changes in genomic network structure can then be used by neurobiologists to identify potential targets for further experimental investigation.

Introduction

In biological systems, networks are rarely static. For instance, as a stem cell differentiates into a neuron, the regulatory relationships between genes shift dynamically. Modeling these time-varying networks presents a significant statistical challenge known as "two-way sparsity." A robust model must simultaneously account for sparsity within a single time point (most genes do not interact) and sparsity across time (network changes should be parsimonious and smooth).

Snapshots of a time-varying network showing structural changes over time steps t=1, 2, 3.

Existing methods often struggle to balance these constraints, particularly when scaling to high-dimensional genomic data. This paper introduces a Bayesian approach that handles high-dimensional single-cell RNA sequencing data while maintaining this delicate balance between sparsity and temporal smoothness.

Methodology

The authors propose a node-wise regression framework. By modeling each node's connectivity independently, the approach becomes highly parallelizable across computing clusters. The core innovation lies in a novel prior structure that decouples marginal sparsity from temporal dependence.

Specifically, the model uses Laplace marginal distributions for regression coefficients to enforce sparsity within each time step, effectively implementing a Bayesian Lasso. To link these coefficients across time, a Gaussian copula is employed. This allows the sparsity parameter (lambda) and the temporal smoothness parameter (rho) to be tuned independently, providing greater flexibility than standard multivariate distributions. Inference is performed via an efficient Gibbs sampler that incorporates a "reverse-exponential prior" to discourage sudden, massive jumps in network structure.

The proposed prior structure linking regression coefficients across time points using a Gaussian copula.

Results and Biological Validation

The model was evaluated using both simulated data and single-cell transcriptome data from human neural development (Nowakowski et al., 2017). In simulations, the model achieved high accuracy, with an AUC of 0.9 or higher for network recovery.

When applied to real-world genomic data involving over 22,000 genes, the model successfully identified biologically relevant shifts in gene connectivity. For example, the stem cell marker MOXD1 exhibited high early connectivity that diminished over time, while the mature neuron marker SATB2 gained connections as cells matured. Notably, the model identified similar key regulatory genes as SCENIC (a popular static method) but added the crucial dimension of when those connections actually matter. Furthermore, the time-varying nature of the model provided resilience against "dropouts"—a common noise source in single-cell sequencing—by interpolating through missing values using the temporal prior.

Conclusion

By leveraging copulas to separate sparsity and smoothness, this framework offers a scalable solution for inferring dynamic networks in high-dimensional settings. While the current model identifies co-expression rather than direct physical regulation, it provides a powerful tool for neurobiologists to identify potential targets for further experimental investigation. Future work may incorporate physical binding data (e.g., ChIP-seq) and explicit zero-inflation modeling to further refine these insights.

Summary updated

Abstract

We propose a novel way of modelling time-varying networks by inducing two-way sparsity on local models of node connectivity. This two-way sparsity separately promotes sparsity across time and sparsity across variables (within time). Separation of these two types of sparsity is achieved through a novel prior structure which draws on ideas from the Bayesian lasso and from copula modelling. We provide an efficient implementation of the proposed model via a Gibbs sampler, and we apply the model to data from neural development. In doing so, we demonstrate that the proposed model is able to identify changes in genomic network structure that match current biological knowledge. Such changes in genomic network structure can then be used by neurobiologists to identify potential targets for further experimental investigation.
BibTeX
@article{bartlett2021two,
  author = {Thomas E. Bartlett and Ioannis Kosmidis and Ricardo Silva},
  title = {{Two-way sparsity for time-varying networks with applications in genomics}},
  journal = {The Annals of Applied Statistics},
  year = {2021},
  doi = {10.1214/20-aoas1416},
}
Figure