🧬 biology

State- versus Reaction-Based Information Processing in Biochemical Networks

This paper demonstrates that the systematic information loss observed in biochemical networks under the Linear-Noise Approximation arises from its reliance on state-based trajectory descriptions, and proposes a reaction-based Gaussian framework that accurately captures information transfer by preserving the sequence of reaction events.

Original authors: Anne-Lena Moor, Age Tjalma, Manuel Reinhardt, Pieter Rein ten Wolde, Christoph Zechner

Published 2026-06-18

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Anne-Lena Moor, Age Tjalma, Manuel Reinhardt, Pieter Rein ten Wolde, Christoph Zechner

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine a busy kitchen where a head chef (the input signal) is trying to tell a line cook (the output system) exactly how to prepare a dish. The chef doesn't speak; instead, they communicate by dropping ingredients into a pot. Every time an ingredient drops, it's a "reaction event."

For a long time, scientists studying these biological "kitchens" (biochemical networks) used a simplified way to measure how much information the cook actually received. They looked at the total amount of food in the pot over time. If the pot got heavier, they assumed the cook knew the chef was busy. This method is called the State-Based approach. It treats the pot like a smooth, continuous liquid where you only see the final level of the water.

However, the authors of this paper argue that this smooth view is missing the most important part of the story. They propose a Reaction-Based approach, which is like watching a high-speed camera recording every single ingredient dropping into the pot. You can see exactly what fell, when it fell, and in what order.

Here is the breakdown of their discovery using simple analogies:

1. The Two Ways to Watch the Kitchen

The State-Based View (The "Blurred" Camera):
Imagine you are watching the pot from far away. You see the water level rise and fall. If two different ingredients (say, salt and pepper) are dropped in, you just see the water level go up. You can't tell which one caused the rise. In this view, the "noise" (the random dropping of ingredients) is mixed together into one big, blurry fluctuation.
- The Problem: If the chef drops salt to signal "spicy" and pepper to signal "sweet," but you only see the water level rise, you lose the specific message. You can't tell if the pot is spicy or sweet, just that it's "fuller." This leads to a loss of information.
The Reaction-Based View (The "High-Speed" Camera):
Now, imagine you have a camera that records every single drop. You see a blue drop (salt) followed by a red drop (pepper). Even if the water level looks the same as in the blurred view, you now know the exact sequence of events.
- The Advantage: This view captures the sequence. It knows that the "spicy" signal happened before the "sweet" signal. It preserves the specific details that the blurred view threw away.

2. Why the Old Math Was Wrong (Even for Big Systems)

Scientists used to think that if a kitchen was huge (with millions of ingredients), the "blurred" view would be accurate enough. They thought the small details didn't matter when the numbers were big.

The paper shows this is not true. Even in a massive system with huge numbers of molecules, the "blurred" view (State-Based) still throws away crucial information. It's like trying to understand a complex song by only listening to the volume of the music, ignoring the melody and the lyrics. No matter how loud the music gets, you still miss the song.

The authors found that the standard math used for years (called the Linear-Noise Approximation) is built on the "blurred" view. Therefore, it systematically underestimates how much information cells are actually exchanging.

3. The "Data-Processing" Analogy

The paper uses a concept called the "Data-Processing Inequality." Think of it like this:

Step 1: The chef drops a specific ingredient (Reaction 1). This contains the secret message.
Step 2: Another ingredient is dropped randomly (Reaction 2) just to fill space.
The Blur: If you mix these two drops together into a single "water level" measurement, you have mixed the secret message with random noise. You can never separate them again.
The Result: The more you mix things together (summarizing reactions), the less information you have. The "blurred" view does this mixing automatically. The "high-speed" view keeps them separate, preserving the message.

4. Real-World Examples in Cells

The authors tested this on two common biological setups:

Example A (The Filter): Imagine a system where a signal passes through a filter. The "blurred" view thinks the filter makes the signal weaker because it mixes the signal with background noise. The "high-speed" view sees that the signal is actually still clear; it just needs to be read in the right way (by looking at the specific reaction, not the total mix).
Example B (The Infinite Signal): In one specific setup, the "blurred" view predicted that the information transfer was infinite (which is impossible and a sign the math broke). The "high-speed" view correctly showed that the information was finite and realistic.

5. How Cells Might "Hack" This

The paper asks: "If the 'blurred' view loses information, how do cells get the full message?"

They look at Ligand-Receptor Binding (how cells sense chemicals).

Old Idea (Berg-Purcell): Cells count how many receptors are "occupied" on average over time. This is like the "blurred" view. It works, but it's not perfect.
New Idea (Maximum Likelihood): Cells can actually count the specific events of binding and unbinding. This is like the "high-speed" view.
The Trick: The paper suggests that cells might use complex, multi-step processes (like a receptor that changes shape several times before releasing a chemical) to "clean up" the noise. By doing this, the cell can effectively turn a "blurred" signal into a "high-speed" signal, allowing it to access the full amount of information hidden in the reaction events.

The Bottom Line

The paper concludes that the difference between a cell's complex reality and our simple math models isn't just about the number of molecules (discrete vs. continuous). The real difference is how we choose to look at the data.

If we look at the total amount of stuff (State-Based), we lose the story.
If we look at the sequence of events (Reaction-Based), we keep the story.

The authors show that by switching our mathematical lens to focus on the sequence of reactions, we can accurately predict how much information cells are actually sharing, fixing a long-standing error in how we understand biological communication.

Technical Summary: State- versus Reaction-Based Information Processing in Biochemical Networks

Problem Statement
Accurately quantifying information transmission in biochemical systems is essential for understanding cellular signaling. The standard metric for this is the trajectory mutual information (or path mutual information rate), which captures not only instantaneous input-output mapping but also temporal correlations. While the Linear-Noise Approximation (LNA) combined with Gaussian channel theory is widely used to calculate these rates—particularly under the assumption of large copy numbers—recent observations indicate qualitative discrepancies between LNA results and exact Markov jump process formalisms. Specifically, the Gaussian approach can significantly underestimate information rates or predict qualitatively different behaviors, even in regimes where copy numbers are large and instantaneous mutual information is correctly predicted. The central problem addressed is understanding the origin of these discrepancies and determining how information is encoded in biochemical trajectories.

Methodology
The authors analyze well-mixed reaction systems of size $\Omega$ involving $M$ species and $K$ reaction channels, modeled as discrete Markov jump processes. They employ the Linear-Noise Approximation (LNA) to derive continuous stochastic differential equations for fluctuations around the macroscopic mean.

The core methodological innovation is the distinction between two definitions of trajectories within the Gaussian formalism:

State-Based Trajectories: These are defined by the effective copy number (or concentration) fluctuations of the chemical species. In this view, the noise vector is $M$ -dimensional, derived by projecting the $K$ -dimensional reaction noise vector onto the species space via the stoichiometry matrix. This effectively sums distinct reaction channels into a single noise term for each species.
Reaction-Based Trajectories: These are defined by the sequence of individual reaction events. In this view, the trajectory is described by the $K$ -dimensional reaction noise vector, preserving the identity of individual reaction channels.

The authors derive explicit expressions for the trajectory mutual information rate for both definitions. For state-based trajectories, they utilize standard Kalman-Bucy filtering equations. For reaction-based trajectories, they adapt filtering theory to calculate the required conditional variances, noting that the standard Kalman-Bucy filter does not directly apply due to the different noise structure.

Key Contributions and Results

Source of Information Loss: The paper demonstrates that the projection from reaction-based increments to state-based increments is irreversible. In the state-based formalism, distinct reaction channels are summarized into effective copy number fluctuations. If a subset of reactions carries information about the input while others act as uninformative noise, the state-based approach cannot distinguish between them. Consequently, the uninformative noise dilutes the signal, leading to a systematic loss of information. This is formally explained via the data-processing inequality: summarizing reactions (post-processing) without additional signal information can only decrease mutual information.
Quantitative Discrepancies:
- Motif Analysis: The authors analyze two common reaction motifs. In Motif (a), the reaction-based formalism yields a higher information rate because it explicitly tracks the specific reaction carrying the signal, whereas the state-based approach conflates it with a non-informative reaction, attenuating the signal.
- Divergence vs. Finiteness: In Motif (b), the state-based formalism predicts an infinite information rate due to instantaneous anti-correlation of species noise terms (a mathematical artifact of the projection). In contrast, the reaction-based formalism correctly yields a finite information rate by tracking individual reaction events.
Connection to Sensing Mechanisms: The paper links these findings to established biological sensing strategies:
- Berg-Purcell (BP) Sensing: Relies on time-averaged receptor occupancy, which corresponds to the state-based trajectory description.
- Maximum-Likelihood (ML) Sensing: Relies on the duration of unbound states (binding events), which corresponds to the reaction-based trajectory description.
- The authors show that the ML sensing error is lower than the BP error by a factor of 2, consistent with the reaction-based information rate being higher than the state-based rate.
Accessibility of Reaction-Based Information: The paper investigates whether cells can access the higher information rates predicted by the reaction-based formalism. By analyzing a multi-state receptor model (where unbinding involves a sequence of steps), they demonstrate that increasing the number of intermediate states narrows the distribution of unbinding times. This reduces the noise contribution from unbinding events, allowing the state-based mutual information rate to approach the reaction-based limit. This suggests that non-equilibrium reaction cycles can effectively "read out" information encoded in specific reaction events.

Significance and Claims
The paper claims that the fundamental difference between the widely used Gaussian approximation (LNA) and the exact Markov jump process is not the discreteness of molecules, but rather the encoding of information in reaction events versus concentrations.

Systematic Underestimation: The commonly used LNA-based Gaussian mutual information is consistent with a state-based trajectory notion, which inherently causes a systematic loss of information independent of system size.
Corrective Framework: An alternative, reaction-based variant of the Gaussian mutual information is proposed, which prevents this loss and accurately predicts information transmission for linear chemical systems at large copy numbers.
Biological Implication: The findings suggest that biological systems may utilize specific network motifs (such as multi-step reaction cycles) to mitigate the information loss inherent in state-based descriptions, thereby accessing the higher information capacity stored in the sequence of reaction events. This connects stochastic thermodynamics concepts (where coarse-graining leads to entropy underestimation) with information theory in biochemical networks.