Quantifying structural uncertainty in chemical reaction network inference
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "Quantifying structural uncertainty in chemical reaction network inference".
Jane: Here's the gist of the paper. Chemical reaction network (CRN) inference aims to identify unknown reactions between species from observed concentrations over time, and this work explores how effective sparse regularisation methods are for quantifying structural uncertainty in CRN inference Reaction networks are frequently used to describe biological systems in fields ranging from cell biology to ecology Chemical reaction networks (CRNs) are a useful framework for exploring biochemical systems in particular, where interactions between the chemical species are described as reactions Parameter estimation is often accompanied by uncertainty and identifiability analyses, e.g., using profile likelihood methods
4: or Bayesian methods
5, 6: However, these analyses mostly pertain to uncertainty that arises from parameter estimates; less attention has been paid to the uncertainty stemming from the specified model structure This does not reflect well the fact that one often does not have complete knowledge of the underlying species interactions It is common wisdom that one cannot trust parameter estimates without uncertainty quantification; the same should be true for network inference This motivates the problem of CRN inference which we tackle in this work: given a family of candidate CRN structures, we seek to find a subset of CRNs that are plausible with respect to observed time-series data Structural uncertainty is represented by a collection of CRNs, each weighted by a model probability These CRNs are obtained from a collection of locally optimal solutions to a sparse regularisation problem The approach does not guarantee that all plausible CRNs will be found via sparse regularisation; we find that the coverage of plausible CRNs depends on the regularisation induced by the penalty function Furthermore, we demonstrate that CRN coverage can be improved by a recombination strategy Structural uncertainty is represented by a collection of CRNs, each weighted by a model probability Mapping Estimates to Structures In the pruning stage, for each estimate (kˆ,σˆ 2) ≡ θˆ ∈ Θˆ, we determine which reactions contribute significantly to the system dynamics induced by θˆ The contribution of reaction r is quantified by g(r; kˆ) = Z tN 0 hat kr Y S s=1 x˜s(t; kˆ) m− rs dt Given that kˆ was obtained with sparse optimisation, we prune reactions with negligible values of g(r; kˆ), thus mapping θˆ to the CRN consisting of the remaining reactions Mapping the collection of estimates Θˆ results in a base set of CRNs, which we denote as Rbase(Θˆ) However, since it is unlikely that all local minima of (4) are found during parameter inference, we expect Rbase(Θˆ) to be missing some CRNs that can plausibly explain the data The recombination stage aims to recover such CRNs by combining CRNs from Rbase(Θˆ) Evaluation and Interpretation of Results The posterior distribution over CRNs is approximated by p(RD) ≈ p(R) exp(−BIC(R)/2) for each R ∈ R(Θˆ) The comparison of penalty functions will be a key theme in our results We find that the L1 penalty function results in less parsimonious CRNs compared to the nonconvex penalty functions The recombination procedure is instrumental in finding the top CRNs Key Findings on Uncertainty We find that ignoring structural uncertainty leads to unreliable predictions (Figure 2) Structural ambiguities elicited from a network posterior can guide how a biological system should be perturbed in future experiments to optimise their utility in narrowing down structural uncertainty The tree representation of a 95% HPD set of CRNs is a novel visualisation of structural uncertainty for CRN inference, allowing structural ambiguities to be highlighted Limitations and Future Directions The approximate posterior probabilities (10) are approximate in two senses, namely due to approximating the model evidence (9) and truncating the exact posterior (6) A future direction of research is to make CRN inference fully Bayesian The second approach requires performing inference jointly over CRN space and their corresponding parameter spaces, typically done with a RJMCMC scheme where a CRN and its rate constants are jointly sampled Supplementary Details The objective function is minimized over log-transformed variables η and ς The scaling factors κr are determined by estimating the order of magnitude of each kr, and we confirmed that our scaling factors are adequate The hierarchical representation corroborates the alternative reaction sets suggested from posterior correlations The L1 penalty results in poorer coverage of the ground-truth CRN and the dynamically equivalent CRNs Candidate Reactions The library of candidate reactions Rall consists of X1 k1 → X2 X3 k1 1 → X2 + X3 k2 1 → X3 X3 k8 2X1 + X2 The library of candidate reactions Rall consists of the following reactions Case Studies In Case study 1, we find that for all penalty functions, the 95% HPD set consists of more than 100 CRNs, indicating a large degree of structural uncertainty In Case study 2, we find that the denitrogenation of pyridine features 7 species, and data is reported for 6 species at 12 time points We find that the L1 penalty function results in less parsimonious CRNs compared to the nonconvex penalty functions When cast as a sparse regularisation problem, CRN inference resembles the problem of variable selection in statistical literature The poor performance cannot be salvaged by simply increasing the penalty hyperparameter, as the largest penalty hyperparameters we used in this work for the L1 penalty resulted in overly simple CRNs during the pruning stage Data Availability The code and data used in this work is available at https://github.com/ysfoo/crn-inference The code and data used in this work is available at https://github.
Paper summary: Tom: So to wrap up, this paper, "Quantifying structural uncertainty in chemical reaction network inference," shows that sparse regularisation methods are used to find a set of candidate reaction networks from data, but they don't give you a complete picture of all possibilities.
Jane: The main point is that you need to quantify this structural uncertainty because ignoring it leads to unreliable predictions, and this paper explores how effective sparse regularisation methods are for quantifying it <ref:2505.15653#pg1>.
Lu: They demonstrate that the recombination procedure is instrumental in finding the top candidate reaction networks, suggesting that combining structures from a base set can improve coverage <ref:2505.15653#pg2>.
Meng: From a deployment view, it suggests that tuning the penalty function is crucial because different choices affect how many plausible networks you actually find in your pruning stage <ref:2505.15653#pg2>.
Tom: And they found that the L1 penalty function gives less parsimonious reaction networks compared to some of the nonconvex functions, and they show that simply increasing the penalty hyperparameter doesn't fix poor coverage during pruning <ref:2505.15653#pg3>.
Jane: The implication for us is that we need a way to handle this ambiguity, maybe using visualizations like the tree representation of the ninety-five percent highest posterior density set of CRNs <ref:2505.15653#pg4>, which helps guide future experiments <ref:2505.15653#pg4>.
Lalam: Basically, when you infer a chemical system, it’s not just finding one answer; it’s about understanding the whole set of plausible structures and how those uncertainties interact with your measurements <ref:2505.15653#pg1>.
Conclusion: Tom: So, we've been looking at this paper about chemical reaction networks and figuring out how to map those observed concentrations back to the actual reactions that could be happening in a system over time.
Jane: Exactly, it’s about taking those observations and not just finding one possible set of reactions, but understanding the whole family of possibilities—the structural uncertainty—that could explain what we see.
Lu: The core idea here is using sparse regularisation to generate these candidate reaction networks, but the authors are focused on how well that process actually covers all the right candidates.
Meng: From a deployment standpoint, that’s important because if you can't tell which structures are plausible, you can't reliably predict what happens next when you run that system in the real world.
Lalam: I found they introduce a recombination strategy to try and fill in those gaps between the initial set of candidate networks they generate.
Tom: That sounds like a way to clean up the output from their main search method, which is pretty smart if it means we're not missing crucial structures during inference.
Jane: But there's a catch, because the L1 penalty function they used actually resulted in fewer reaction networks being considered as plausible compared to other types of penalty functions.
Lu: Right, so the choice of how you punish complexity in your math directly impacts how complete your final set of structures turns out to be.
Meng: And they found that just cranking up the penalty setting doesn't fix that coverage problem; it actually made the networks too simple during their initial search stage.
Tom: So, it’s not a simple knob you can turn up to get better results, which means we have to be careful about how we set up these kinds of sparse problems.
Jane: The authors show that if you ignore this structural uncertainty, your predictions based on the network are going to be unreliable in practice.
Lu: And they suggest using the structure of these uncertainties—like a tree representation—as a way to guide future experiments to narrow down what's actually plausible.
Meng: That visualization of uncertainty sounds like something that could be really useful for debugging models when we’re trying to understand why a prediction failed at scale.
Tom: It means understanding the ambiguity isn't just an academic exercise; it has direct implications for how we design experiments or how we trust the model outputs.
Yong See Foo, Adriana Zanca, Jennifer A. Flegg, Ivo Siekmann
School of Mathematics and Statistics, University of Melbourne · ARC Centre of Excellence for the Mathematical Analysis of Cellular Systems, University of Melbourne · School of Computer Science and Mathematics, Liverpool John Moores University
stat.ME, q-bio.QM
Submitted: 2025-05-21
Updated: 2026-10-31
Comments: 41 pages, 12 figures
Code: https://github.com/ysfoo/crn-inference
License: http://creativecommons.org/licenses/by-nc-sa/4.0/
Importance score: 75/100
The gist: Chemical reaction network (CRN) inference aims to identify unknown reactions between species from observed concentrations over time…
Key concepts
- Chemical Reaction Network (CRN) Inference
- This is the process of identifying unknown chemical reactions within a system based on observed concentrations over time. It helps map out how different chemical species interact through reactions, which is crucial for understanding biological systems.
- Structural Uncertainty
- This refers to the ambiguity about the actual structure of the reaction network—which specific reactions are present or absent. Because we don't always know the complete model, this uncertainty must be quantified alongside parameter uncertainty.
- Sparse Regularisation
- This is a mathematical technique used to find simpler models by penalizing overly complex structures. In CRN inference, it helps select only the most significant reactions that explain the observed data, making the search for plausible networks more manageable.
Terminology
Summary
Chemical reaction network (CRN) inference aims to identify unknown reactions between species from observed concentrations over time, and this work explores how effective sparse regularisation methods are for quantifying structural uncertainty in CRN inference
Introduction to the Problem
Reaction networks are frequently used to describe biological systems in fields ranging from cell biology to ecology Chemical reaction networks (CRNs) are a useful framework for exploring biochemical systems in particular, where interactions between the chemical species are described as reactions Parameter estimation is often accompanied by uncertainty and identifiability analyses, e.g., using profile likelihood methods [4] or Bayesian methods [5, 6] However, these analyses mostly pertain to uncertainty that arises from parameter estimates; less attention has been paid to the uncertainty stemming from the specified model structure This does not reflect well the fact that one often does not have complete knowledge of the underlying species interactions It is common wisdom that one cannot trust parameter estimates without uncertainty quantification; the same should be true for network inference This motivates the problem of CRN inference which we tackle in this work: given a family of candidate CRN structures, we seek to find a subset of CRNs that are plausible with respect to observed time-series data
Methodology for Uncertainty Quantification
Structural uncertainty is represented by a collection of CRNs, each weighted by a model probability These CRNs are obtained from a collection of locally optimal solutions to a sparse regularisation problem The approach does not guarantee that all plausible CRNs will be found via sparse regularisation; we find that the coverage of plausible CRNs depends on the regularisation induced by the penalty function Furthermore, we demonstrate that CRN coverage can be improved by a recombination strategy Structural uncertainty is represented by a collection of CRNs, each weighted by a model probability
Mapping Estimates to Structures
In the pruning stage, for each estimate (kˆ,σˆ 2) ≡ θˆ ∈ Θˆ, we determine which reactions contribute significantly to the system dynamics induced by θˆ The contribution of reaction r is quantified by g(r; kˆ) = Z tN 0
hat kr Y S s=1 x˜s(t; kˆ) m− rs dt Given that kˆ was obtained with sparse optimisation, we prune reactions with negligible values of g(r; kˆ), thus mapping θˆ to the CRN consisting of the remaining reactions Mapping the collection of estimates Θˆ results in a base set of CRNs, which we denote as Rbase(Θˆ) However, since it is unlikely that all local minima of (4) are found during parameter inference, we expect Rbase(Θˆ) to be missing some CRNs that can plausibly explain the data The recombination stage aims to recover such CRNs by combining CRNs from Rbase(Θˆ)
Evaluation and Interpretation of Results
The posterior distribution over CRNs is approximated by p(RD) ≈ p(R) exp(−BIC(R)/2) for each R ∈ R(Θˆ) The comparison of penalty functions will be a key theme in our results We find that the L1 penalty function results in less parsimonious CRNs compared to the nonconvex penalty functions The recombination procedure is instrumental in finding the top CRNs
Key Findings on Uncertainty
We find that ignoring structural uncertainty leads to unreliable predictions (Figure 2) Structural ambiguities elicited from a network posterior can guide how a biological system should be perturbed in future experiments to optimise their utility in narrowing down structural uncertainty The tree representation of a 95% HPD set of CRNs is a novel visualisation of structural uncertainty for CRN inference, allowing structural ambiguities to be highlighted
Limitations and Future Directions
The approximate posterior probabilities (10) are approximate in two senses, namely due to approximating the model evidence (9) and truncating the exact posterior (6) A future direction of research is to make CRN inference fully Bayesian The second approach requires performing inference jointly over CRN space and their corresponding parameter spaces, typically done with a RJMCMC scheme where a CRN and its rate constants are jointly sampled
Supplementary Details
The objective function is minimized over log-transformed variables η and ς The scaling factors κr are determined by estimating the order of magnitude of each kr, and we confirmed that our scaling factors are adequate The hierarchical representation corroborates the alternative reaction sets suggested from posterior correlations The L1 penalty results in poorer coverage of the ground-truth CRN and the dynamically equivalent CRNs
Candidate Reactions
The library of candidate reactions Rall consists of X1 k1 → X2 X3 k1 1 → X2 + X3 k2 1 → X3 X3 k8 2X1 + X2 The library of candidate reactions Rall consists of the following reactions
Case Studies
In Case study 1, we find that for all penalty functions, the 95% HPD set consists of more than 100 CRNs, indicating a large degree of structural uncertainty In Case study 2, we find that the denitrogenation of pyridine features 7 species, and data is reported for 6 species at 12 time points
Conclusion on Sparse Regularisation
We find that the L1 penalty function results in less parsimonious CRNs compared to the nonconvex penalty functions When cast as a sparse regularisation problem, CRN inference resembles the problem of variable selection in statistical literature The poor performance cannot be salvaged by simply increasing the penalty hyperparameter, as the largest penalty hyperparameters we used in this work for the L1 penalty resulted in overly simple CRNs during the pruning stage
Data Availability
The code and data used in this work is available at https://github.com/ysfoo/crn-inference The code and data used in this work is available at https://github. p. 0 This representation translates into alternative reaction
pathways suggested by the available data, thus guiding the efforts of future experimental design. Keywords equation discovery · model selection · model uncertainty · network inference · reaction networks
1 Introduction
Reaction networks are frequently used to describe biological systems in fields ranging from cell biology to ecology [1–3].Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Model uncertainty quantification in chemical reaction network inference allows AI systems to identify alternative reaction pathways suggested by available data, guiding future experimental design rather than relying on a single inferred model structure.
-
The system can generate a hierarchical representation of structural ambiguities, which provides higher-order information about structural uncertainty that is not immediately obvious from posterior correlations alone.
-
By employing recombination strategies based on exchange pairs between plausible CRNs, the AI system can recover CRNs with high posterior probabilities that were missed by continuous optimization methods.
-
The AI system can perform rigorous model selection by comparing the coverage of plausible CRNs across different sparse regularization penalty functions, identifying which penalty function results in more parsimonious and accurate models for a given data regime.
-
The system can be used to assess the reliability of predictions by comparing trajectory reconstruction errors using the posterior mode CRN against novel initial states, flagging situations where structural uncertainty leads to overconfident and unreliable inferences.
Abstract
Dynamical systems in biology are complex, and one often does not have comprehensive knowledge about the interactions involved. Chemical reaction network (CRN) inference aims to identify, from observing species concentrations over time, the unknown reactions between the species. Existing approaches such as sparse regularisation largely focus on identifying a single, most likely CRN, without addressing uncertainty about the network structure. However, it is important to quantify structural uncertainty to have confidence in our inference and predictions. In this work, we explore how effective sparse regularisation methods are for quantifying structural uncertainty. Locally optimal solutions to sparse regularisation are mapped to CRN structures; however, it is unclear whether this approach encompasses all plausible CRNs. We find that inducing sparsity with nonconvex penalty functions results in better coverage of the plausible CRNs compared to the popular lasso regularisation. To validate our approach, we apply our methods to real-world data examples, and are able to simultaneously recover reactions proposed across multiple literature sources for a reaction system. Our emphasis on network-level probabilities enables a novel, hierarchical representation of structural ambiguities in the space of CRNs. This representation translates into alternative reaction pathways suggested by the available data, thus guiding the efforts of future experimental design.
Related papers
- Learning Robust Treatment Rules for Censored Data
- Neural Estimation of Summary Statistics for Simulation-Based Inference with Non-Asymptotic Guarantees
- Spectral divide-and-conquer MCMC for long stationary time series
- Improving the Efficiency of Subgroup Analysis in Randomized Controlled Trials with TMLE
- Personalized Federated Vector Autoregression with Personalization Diversity
- Doubly Robust and Efficient Calibration of Prediction Sets for Right-Censored Time-to-Event Outcomes