A Data-Centric Diagnostic Framework for Experimental Design and Mechanistic Modeling: Application to mRNA Production

Abstract In this study, we present a data-centric framework for the integrated investigation of experimental behavior and mechanistic model performance in complex biochemical systems. The framework is demonstrated on the in vitro transcription (IVT) reaction for mRNA production. Rather than analyzing experimental data and mechanistic models independently, the framework links experimental-response surrogate modeling, a surrogate model of error characterizing model–experiment discrepancy, and mechanistic parameter sensitivity within a unified workflow. Gaussian Process Regression (GPR) and Global Sensitivity Analysis (GSA) are first used to interrogate experimental data, identify statistically anomalous observations, and characterize the process variables and interactions governing the response. Predictions from the resulting experimental-response surrogate are then compared with those of an existing mechanistic model, and the resulting discrepancy is represented through a surrogate model of error and subjected to GSA. Finally, the process-level drivers of this discrepancy are cross-compared with sensitivities of the mechanistic kinetic parameters, thereby linking experimental and mechanistic information to prioritize subsequent investigation. The experimental-response surrogate achieved an R2 of 0.90, with MgOAc and NTP concentrations dominating variation in mRNA yield and incubation time contributing principally through interactions. The surrogate model of error achieved an R2 of 0.86 and identified NTP, MgOAc, T7-RNA polymerase, and selected interactions as the principal contributors to the discrepancy between mechanistic and experimental-surrogate predictions. Cross-comparison with mechanistic parameter sensitivities subsequently prioritized the Mg degradation rate constant (kMg), DNA equilibrium constant (KDNA), Mg saturation constant (k1), and transcription rate constant (kapp), together with their associated model components, for further investigation. The resulting framework provides a systematic basis for directing subsequent repeat or targeted experimentation, parameter recalibration, and structural model refinement.