Effects of Model Reduction on Coherence and Information Transfer in Stochastic Biochemical Systems

Juan David Marmolejo Lozano, Nikola Popović, Ramon Grima · 2025

Read Effects of Model Reduction on Coherence and Information Transfer in Stochastic Biochemical Systems by Juan David Marmolejo Lozano, Nikola Popović, Ramon Grima (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-27600.txt.

Introduction

Juan David Marmolejo Lozano, Nikola Popovic, and Ramon Grima warn in 2025 that simplified cell models can mislead on signal flow. Cells run networks of chemical steps, and counts of each kind jitter at random. Two gauges track that jitter: coherence, which shows if two counts move in step by frequency, and the mutual information rate, which shows how much one count tells about another.

Reduced models that split fast and slow steps match plain count stats well. Yet they often err badly on coherence at mid and high frequencies. Those errors then skew the information rate. The flaw stems from network shape and from which steps were cut. Care is needed when small models stand in for living networks.

Catalog record

CREATORJuan David Marmolejo Lozano, Nikola Popović, Ramon Grima
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE102 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-27600

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Effects of Model Reduction on Coherence and Information Transfer in Stochastic Biochemical Systems Juan David Marmolejo Lozano, Nikola Popović, Ramon Grima Abstract Simplified stochastic models are widely used in the study of frequency-resolved noise propagation in biochemical reaction networks, a common measure being the coherence between random fluctuations in molecule number trajectories. Such models have also found widespread application in the quantification of how information is transmitted in reaction networks via the mutual information (MI) rate. A common assumption is that, under timescale separation, estimates for the coherence and MI rate obtained from simplified (reduced) models closely approximate those in the underlying full models. Here, we challenge that assumption by showing that, while reduced models can faithfully reproduce low-order statistics of molecular counts, they frequently incur substantial discrepancies in the coherence spectrum, especially at intermediate and high frequencies. These errors, in turn, lead to significant inaccuracies in the resulting estimates for the MI rates. We show that the observed discrepancies are due to the interplay between the structure of the underlying reaction networks, the specific model reduction method that is applied, and the asymptotic limits relating the full and the reduced models. We illustrate our results in canonical models of enzyme catalysis and gene expression, highlighting practical implications for quantifying information flow in cells. I Introduction Realistic models of stochastic biochemical systems are typically difficult to study analytically, or even through stochastic simulation, due to the large number of interacting species and the myriad interactions involved. A much more common approach relies on the construction and study of reduced models that capture the essential dynamics of the underlying “full” systems, but that involve reduced numbers of interacting species or simplified reaction kinetics. Several methodologies have been devised to achieve model reduction, of which three are commonly used: (i) reduction through separation of timescales Thomas et al. 2012a ; Kim et al. 2014 ; Song et al. 2021 ; Popović et al. 2016 ; Kang et al. 2019 ; Thomas et al. 2014 ; (ii) reduction through abundance-scale separation Jahnke 2011 ; Smith et al. 2015 ; Hellander and Lötstedt 2007 ; Cao and Grima 2018 ; Jia and Grima 2024 ; and (iii) reduction by means of effective time-delayed reactions Barrio et al. 2013 ; Leier et al. 2014 ; Zhang and Zhou 2019 ; Szavits-Nossan and Grima 2024 . Some of these methodologies are equivalent to minimizing the Kullback-Leibler divergence between the full and the reduced models, defined on the space of Monte Carlo simulation trajectories Öcal et al. 2023 . Of the three types of model reduction, the most widely used are methods based on timescale separation; henceforth, we thus exclusively focus on that type. Reduced models of timescale separation type can be formally proven to estimate accurately molecule number statistics, such as moments and distributions, in subregions of parameter space. A prominent example is the slow-scale Linear-Noise Approximation (ssLNA), the unique reduced form of the standard Linear-Noise Approximation (LNA) of a complex biochemical system under timescale separation conditions, which can be rigorously derived using projection operator techniques Thomas et al. 2012a . Specifically, the ssLNA leads to an accurate reduced model when perturbations about the steady-state concentrations of a subset of the species — the slow species — decay on a much slower timescale than those about the steady-state concentrations of the remaining (fast) species; thus, the ssLNA is the stochastic equivalent of the deterministic quasi-steady-state approximation (QSSA). The large-molecule-number limits of several common stochastic reduction methods which assume that the set of governing reactions can be grouped into subsets of slow and fast reactions Cao et al. 2005 ; Goutsias 2005 ; Rao and Arkin 2003 can be shown to be special cases of the ssLNA Thomas et al. 2012a . The accuracy of the ssLNA in calculating the first and second moments of molecular counts, and their power spectra, has been verified by comparison with stochastic simulation via Gillespie’s exact algorithm Gillespie 1977 for a large variety of reaction networks; examples include reactions catalyzed by enzymes, genetic feedback loops, and biologically detailed models of stochastic gene expression Thomas et al. 2012b ; Ma et al. 2024 . An alternative heuristic type of model reduction which is also commonly found in the literature Elf and Ehrenberg 2003 ; Ziv et al. 2007 ; Rodríguez Martínez et al. 2010 assumes the form of the probabilities for the effective reactions by comparison with the well-known effective macroscopic rate laws, such as in the assumption of Hill-type propensities in the Chemical Master Equation (CME). The heuristic CME and the associated heuristic LNA, derived from such heuristic propensities, are not generally accurate when timescale separation holds Thomas et al. 2012b ; Holehouse and Grima 2019 . Reduced models have also been used to calculate the mutual information (MI) rate between molecule number trajectories of a pair of species in a biochemical reaction network Tostevin and Ten Wolde 2009 . Note that the MI rate is the time-averaged mutual information between a pair of stationary stochastic processes, which is different from instantaneous mutual information; for biochemical networks, the former is more relevant, as the latter cannot correctly quantify the amount of information transmitted per unit time Tostevin and Ten Wolde 2010 ; Meijers et al. 2021 ; Fan and Hilfinger 2024 . A related concept is the coherence between molecule number trajectories of a pair of species, which measures how strongly the two trajectories are related at each frequency — within the linear-noise or Gaussian approximation, the coherence completely determines the MI rate via a simple integral formula Munakata and Kamiyabu 2006 ; Tostevin and Ten Wolde 2009 ; Tostevin and Ten Wolde 2010 ; Komaee 2020 . For non-Gaussian fluctuations, as is typically the case, that formula, while not exact, is still useful as a lower bound for the true MI rate Mitra and Stark 2001 . However, the relationship between the coherence and the MI rate predicted by full and reduced models has not been significantly explored, in contrast to the relationship between the moments of the molecular counts, as previously discussed. Since expressions for these quantities have been calculated exclusively for very simple (reduced) models of biological systems, it has in effect been implicitly assumed that the coherence and information transfer measures associated to the reduced models converge to those in the underlying full models in some kinetic parameter limit. For instance, the study of the MI rate between species X and Y interacting via the reaction scheme \emptyset\rightarrow X\rightarrow\emptyset,~X\rightarrow X+Y,~Y\rightarrow\emptyset in Tostevin and Ten Wolde 2009 can be interpreted as a simple model for the transcription of mRNA X and its subsequent translation into protein Y ; however, clearly that is a reduced model, as many of the biologically important intermediate steps — such as gene state switching, nuclear export, and binding of mRNA to ribosomes — are omitted with the implicit understanding that these details are not important to the calculation of the MI rate, provided the intermediate steps are very fast. The ssLNA lends itself perfectly to the task of understanding the relationship between the coherence and the MI rate in complex biochemical systems and their corresponding reduced models, as it allows for the straightforward

Read the complete text