Thermodynamic Guardrails: Real-Time Monitoring of Physical Consistency in Stochastic Biochemical Models

Jaeyoung Choi

Keywords: Mathematical Biology, Biophysics, QSSA, Model Reduction, Bond Graphs, Michaelis-Menten Kinetics, Thermodynamic Validity, Model Error Detection, Autonomous Scientific Discovery

TL;DR: This paper establishes a proof-of concept for thermodynamics-based automatic model error detection in a stochastic enzyme reaction model, achieved via bond graph modeling.

Abstract: The stochastic total quasi-steady-state approximation (stQSSA) is a vital model reduction technique for managing the computational complexity of multiscale biochemical networks. However, its validity is conditional, and it is known to fail in regimes of tight molecular binding and near-equimolar species concentrations \cite{kim2014validity, song2021universally}. This work introduces a thermodynamic guardrail, based on bond graph formalism, as a robust, first-principles-based diagnostic tool for detecting when the stQSSA produces a physically implausible state. Rooted in the Second Law of Thermodynamics, the guardrail identifies the stQSSA's failure by detecting negative power flow—the product of chemical affinity and flux—which signifies a thermodynamic violation. While the diagnosis is successful, the subsequent attempt to switch to a full Stochastic Simulation Algorithm (SSA) reveals a critical challenge: Even with the error detected, naive re-initialization is insufficient to correct the integrated error accumulated by the stQSSA, suggesting other, effective hand-off methods must be employed. This work validates thermodynamic monitoring as a crucial diagnostic for autonomous scientific agents, providing a "physical conscience" to prevent the propagation of erroneous results, while simultaneously exposing the non-trivial problem of state hand-off for the development of truly robust, self-correcting models.

Submission Number: 96