Deep Learning Reaction Network: a machine learning framework for modeling time resolved data

Abstract

Model-based analysis is essential for extracting information about chemical reaction kinetics in full detail from time-resolved data sets. Such analysis combines experimental hypotheses of the process with mathematical models related to the systems physical mechanisms. This combination can provide a concise description of complex system dynamics and extrapolate kinetic model parameters, such as kinetic pathways, time constants, and species amplitudes. However, the process leading to the final kinetic model requires several intermediate steps in which different assumptions and models are tested, even using different experimental data sets. This approach requires considerable experience in modeling and data comprehension, as poor decisions at any stage of time-resolved data analysis (such as time-resolved spectra and agarose gel electrophoresis) can lead to an incorrect or incomplete kinetic model, resulting in inaccurate model parameters and amplitudes. The Deep Learning Reaction Network (DLRN) can rapidly provide a kinetic reaction network, time constants, and amplitude for the system, with comparable performance and, in part, even better than a classical fitting analysis. Additionally, DLRN works in scenarios in which the initial state is a non-emitting dark state and for multiple timescales. The utility of DLRN is also shown for more than one 2D system, as it performed well for both spectral and time-resolved agarose gel electrophoresis data.