Parametric and non-parametric gradient matching for network inference: a comparison

Leander Dony, Fei He, Michael P. H. StumpfBMC Bioinformatics 2019 · DOI

TL;DR

This paper compares parametric (ODE-based) and non-parametric (GP-based) gradient matching methods for inferring gene regulatory networks from time-series gene expression data. They found that parametric methods can offer comparable or even better performance, especially when prior mechanistic knowledge is available, while non-parametric methods are more computationally efficient and require less prior information.

What Problem Does This Paper Address?

Reverse engineering gene regulatory networks (GRNs) from time-series gene expression data is a significant challenge. This is due to the vast number of potential interactions and the inherent stochasticity of gene expression. While many methods exist, this paper focuses on inferring GRNs using nonlinear ordinary differential equations (ODEs) and their non-parametric counterparts. A key hurdle in ODE-based inference is the computational cost associated with repeatedly solving ODEs. To address this, the authors employ a gradient matching approach, which avoids explicit ODE solving by interpolating data and matching derivatives. The paper aims to compare the performance of parametric (where the ODE structure and parameters are known or inferred) and non-parametric (where the underlying kinetics are unknown) gradient matching methods under various conditions.

Methods

The core methodology revolves around gradient matching. This approach involves:

  1. Data Interpolation: Gene expression time-series data is smoothed and interpolated. The authors primarily use Gaussian Process (GP) regression for this, which allows for analytical estimation of derivatives. They explore both single-output GPs (one per gene) and multi-output GPs (modeling correlations between genes).
  2. Gradient Calculation: The rate of change (derivative) of gene expression is estimated from the interpolated data using the GP's derivative capabilities.
  3. Parameter/Model Optimization:
    • Parametric (ODE-based): For a given ODE model structure, the parameters are optimized by minimizing the squared difference between the ODE's right-hand side (evaluated with predicted gene expression) and the estimated derivative. They consider ODEs with and without prior knowledge of parameters.
  4. Non-parametric (GP-based): This approach models the gene expression dynamics directly using GPs, inferring the network structure without assuming a specific ODE form. The inference focuses on identifying regulatory relationships based on correlations in the GP-derived derivatives.
  5. Model Selection: The Bayesian Information Criterion (BIC) is used to select the best models, and Schwarz weights are calculated to combine inferences from multiple models, effectively performing Bayesian model averaging.
  6. Performance Evaluation: The Area Under the Precision-Recall curve (AUPR) is used as the primary metric to evaluate the accuracy of the inferred GRNs.

Datasets

The authors use simulated gene expression data for evaluation:

  • Deterministic ODE model simulation:
    • 'Non-oscillatory data': Generated from a 5-gene IRMA network in Saccharomyces cerevisiae without noise.
  • 'Oscillatory data': A variation of the 5-gene IRMA network with one edge removed, also deterministic.
  • Simulated stochastic gene expression data: Generated using GeneNetWaver from a 10-gene Saccharomyces cerevisiae network, incorporating both process and observational noise. This dataset is referred to as 'InSilicoSize10-Yeast1_dream4'.

Key Insights and Ideas

  • Parametric vs. Non-parametric Trade-offs: Parametric ODE-based methods can recover mechanistic information (like activation/repression) if prior knowledge is available, potentially leading to better performance. Non-parametric GP-based methods are more flexible when kinetics are unknown and are computationally more efficient.
  • Importance of Prior Knowledge: For parametric ODE inference, having prior knowledge of parameters (e.g., basal transcription and degradation rates) significantly improves performance, especially for inferring directed GRNs with interaction types.
  • Noise Amplification: The gradient matching approach, particularly the derivative estimation from GPs, can be sensitive to noise. This suggests that careful experimental design to produce high-quality, informative data is crucial.
  • Multi-output GPs for Interpolation: For small networks, multi-output GPs can improve interpolation by capturing correlations between genes, leading to better inference results compared to independent single-output GPs. However, this benefit diminishes for larger, noisy datasets due to increased computational complexity.
  • Stochastic Data Challenges: Inferring networks from noisy, stochastic data is significantly harder, leading to an overall decline in performance across all methods. The performance differences between methods become less pronounced, and only a narrow range of scenarios yields meaningful results.
  • Bayesian Model Averaging: Combining inferences from multiple models using BIC weights (Bayesian model averaging) is presented as a powerful strategy to improve the quality of network inference.

Key Results

  • Parametric methods can match or outperform non-parametric methods: Particularly when inferring directed GRNs with interaction types and when prior mechanistic information is available.
  • Non-parametric GP-based methods achieve good performance without prior kinetic information: They often perform comparably to ODE-based methods when only directionality is of interest.
  • Stochastic data significantly reduces inference performance: For noisy data, most methods perform only slightly better than random guessing, highlighting the difficulty of the problem.
  • Data interpolation method matters: Multi-output GPs generally outperform single-output GPs for interpolating oscillatory data in smaller networks.
  • Number of data points: For noise-free data, more data points generally lead to better performance. This trend is less clear for stochastic data, where GP interpolation might struggle to accurately capture dynamics with too many points.
  • Maximum number of parents: The maximum number of parents considered per gene had a limited impact on performance for noise-free data, likely due to BIC regularization.

Limitations and Future Work

  • Computational Cost: While gradient matching reduces the cost compared to explicit ODE solving, it can still be computationally intensive, especially for large networks and multi-output GPs.
  • Sensitivity to Noise: The derivative estimation in gradient matching is sensitive to noise, potentially amplifying fitting errors.
  • Adaptive Gradient Matching: The authors suggest that adaptive gradient matching, which incorporates ODE model structure into GP fitting, could improve robustness to noise and is worth pursuing.
  • Combining Parametric and Non-parametric Methods: A promising future direction is to use non-parametric methods to narrow down the search space of potential networks, followed by ODE-based inference to confirm interactions and extract mechanistic details. This hybrid approach could be more efficient for large-scale inference.
  • Multi-output GPs for Complex Networks: While beneficial for small networks, the reliability of multi-output GPs for interpolating complex, stochastic networks needs further investigation.
  • Bayesian Model Averaging: The authors believe that combining the strengths of various approaches, particularly through Bayesian model averaging, is key to making significant progress in GRN inference.

Summary updated

BibTeX
@article{dony2019parametric,
  author = {Leander Dony and Fei He and Michael P. H. Stumpf},
  title = {{Parametric and non-parametric gradient matching for network inference: a comparison}},
  journal = {BMC Bioinformatics},
  year = {2019},
  doi = {10.1186/s12859-018-2590-7},
}
Figure