Identification of dynamic mass-action biochemical reaction networks using sparse Bayesian methods

Richard M. Jiang, Prashant Singh, Fredrik Wrede, Andreas Hellander, Linda R. Petzold

Abstract

Inferring biochemical reaction networks from noisy, sparse time-series data is a fundamental challenge in systems biology. We present Bayesian Reactive SINDy, a framework that combines a regularized horseshoe prior with a latent variable observational model to identify parsimonious, interpretable models. By avoiding numerical differentiation and providing robust uncertainty estimates, this method outperforms traditional sparse identification techniques in both accuracy and reliability.

Introduction

A central goal in systems biology is the automated inference of biochemical reaction networks—the "wiring diagrams" of life—from molecular concentration data. While methods like Sparse Identification of Nonlinear Dynamics (SINDy) have made significant progress, they face two primary hurdles: sensitivity to noise and a lack of uncertainty quantification. Standard methods often rely on numerical derivatives ($dX/dt$), which become highly biased when data is noisy or sparsely sampled. Furthermore, biological data is frequently "unidentifiable," meaning multiple distinct reaction networks might explain the same observations. Researchers need tools that not only find a likely model but also quantify the range of plausible alternatives.

Methods

The proposed Bayesian Reactive SINDy framework addresses these challenges through two key innovations:

  • Regularized Horseshoe Prior: To achieve a "parsimonious" model (one with the fewest reactions necessary), the authors use a sparse Bayesian prior called the regularized horseshoe. Unlike the standard LASSO (L1) penalty, the horseshoe prior is continuous and more effective at shrinking irrelevant reaction rates toward zero while allowing true reaction rates to remain large and unbiased.
  • Latent Variable Observational Model: Instead of calculating derivatives from noisy data, the authors treat the true system state as a "latent" (hidden) variable governed by an Ordinary Differential Equation (ODE). Observations are modeled as noisy samples from this latent state. This allows the framework to handle sparse sampling, measurement noise, and even "partially observed" systems where some species are unmeasured.
Table 1: A representative library of candidate reactions for a benchmark system. The framework evaluates a large set of "ansatz" reactions to identify the sparse subset (indicated by non-zero rate constants) that best explains the observed data.

The model is implemented using the No-U-Turn Sampler (NUTS) in the probabilistic programming language Stan, enabling efficient sampling of the posterior distribution.

Results and Key Insights

The method was evaluated on synthetic benchmarks designed to mimic real-world biological challenges, including a Lotka-Volterra (predator-prey) system and a complex 5-species prokaryotic auto-regulation network.

  • Robustness to Sparsity: On the Lotka-Volterra system, the Bayesian approach maintained high accuracy at large observation intervals where traditional derivative-based methods failed to recover the correct dynamics.
  • Embracing Ambiguity: In the auto-regulation test, Markov Chain Monte Carlo (MCMC) chains converged to different "modes." This revealed that multiple distinct reaction networks could explain the data equally well. Rather than a failure, this provides crucial information to biologists, suggesting that current data is insufficient to distinguish between certain mechanisms and indicating a need for further experiments.
  • Flexibility in Observation: Because of the latent variable approach, the method can successfully infer a network even if only the sum of two species is observed or if one species is completely unmeasured, provided the system's dynamics are sufficiently coupled.

Discussion

While powerful, the Bayesian approach is computationally more intensive than optimization-based methods, with current scaling limits estimated at around 20 active species. Additionally, the model assumes deterministic mass-action kinetics. Adapting this sparse Bayesian framework to stochastic models (such as the Chemical Master Equation) remains a major open challenge due to the computational difficulty of calculating likelihoods for stochastic processes at the single-cell level.

Summary updated

Abstract

Author summary Reconstructing the correlated reactions that govern a system of biochemical species from observational temporal data is an essential step in understanding many biological systems. To facilitate this process, we propose a robust, data-driven approach based on a sparse Bayesian statistical model. Our approach exploits sparse Bayesian priors and an unbiased observational model to recover a parsimonious, interpretable reaction system from mass-action relations, utilizing very little user input. On a set of simulated test problems, the method demonstrates increased robustness and decreased bias at different levels of measurement variability, while also producing interpretable reaction systems and quantifying uncertainty. As a tool, the approach can be used to flexibly interrogate biological systems while allowing incorporation of potentially uncertain domain knowledge to improve the efficiency and identifiability of the problem.
BibTeX
@article{jiang2022identification,
  author = {Richard M. Jiang and Prashant Singh and Fredrik Wrede and Andreas Hellander and Linda R. Petzold},
  title = {{Identification of dynamic mass-action biochemical reaction networks using sparse Bayesian methods}},
  journal = {PLOS Computational Biology},
  year = {2022},
  doi = {10.1371/journal.pcbi.1009830},
}
Figure