Automated Inference of Chemical Reaction Networks
Systematic methods and tools for managing the complexity
Process Simulation & Optimization - I (T4-9a)
Keywords: system identification, network inference, chemical reaction network, batch process
Methods to automatically reverse engineer chemical or biochemical reaction networks from experimental data are the subject of considerable interest (Crampin et al., 2004). Research into these techniques has largely been confined to the biosciences, but such methods could have significant commercial implications for the chemical and process industries. Software for kinetic fitting - once the network structure is known - is widely available, but the initial inference of this structure is currently time consuming and requires considerable expertise. Hence, methods that can accelerate the discovery of chemical reaction networks from data have substantial commercial and academic potential. In particular, the burgeoning use of high throughput technologies (HTT) in chemical process research and development (e.g. automated workstations for performing experiments in parallel) coupled with improved chemical sensor technology will provide an increase in the quantity and quality of data available during the product development lifecycle. This trend suggests that methods for the reverse engineering of reaction networks from data will become of increasing importance in the future.
One approach has recently been suggested by Burnham et al. (2006) and Searson et al. (2006). This uses time series concentration measurements obtained from batch or fed-batch reactors, and involves the search of a specific class of kinetic rate expression models, guided by statistical indices. This class of models corresponds to unimolecular and bimolecular elementary reaction steps occurring in a well-mixed isothermal reactor. A successful search of the model structures will yield an overall kinetic model that corresponds to a plausible reaction network. However, the existing search method is performed manually and is both time consuming and subjective. Furthermore, the statistical indices used are particularly susceptible to local minima caused by high order variable interactions.
Here, we propose a novel automated approach to efficiently search the model space whilst avoiding these problems. The proposed methodology comprises 3 stages: (1) A branch and bound algorithm – guided by information entropy (IE) and kinetic heuristics instead of the usual statistical indices- is used to incrementally search the space of kinetic rate models in order to construct a small set of candidate models for each chemical species. (2) Robust regression is applied to these sets to detect model terms with a low statistical significance. These are removed to yield refined sets of candidate models. For each species, the refined candidate model with the best IE measure is extracted and combined with the others to form a trial kinetic model of the reaction network. (3) A final screening stage identifies any inconsistencies within the network kinetics and employs heuristics to resolve them. A set of plausible elementary reaction steps comprising the reaction network is then inferred directly from the identified kinetic model.
The procedure is demonstrated, in simulation, with noisy data from a hypothetical reaction network comprising 5 chemical species involved in 4 simultaneous reactions within a batch reactor. It is shown that the automated procedure is able to correctly identify the structure of the reaction network and estimate the reaction rate constants. Some advantages of the method are, firstly, that it is guaranteed - in the worst case scenario - to be as least as efficient as an exhaustive search of the model space. Usually, however, it is between 4-20 times more efficient. Secondly, it has the potential to scale well to larger and more complex systems (e.g. those involving higher order kinetic rate terms corresponding to multi-step reactions involving short-lived intermediates). Finally, the procedure is almost entirely automated; making it a suitable basis for a software tool to aid the discovery of chemical reaction networks from laboratory scale data.
Selected References
Burnham, S.C., Searson, D.P., Willis, M.J. and Wright, A.R., Towards the automated deduction of chemical reaction mechanism, In Proceedings of the 17th CHISA International Congress of Chemical Engineering, Prague, August, 2006.
Crampin, E.J., Schnell, S. and McSharry, P.E., Mathematical and computational techniques to deduce complex biochemical reaction mechanisms. Progress in Biophysics & Molecular Biology, 86, pp. 77–112, 2004.
Searson, D.P., Burnham, S.C., Willis, M.J. and Wright, A.R., Identification of chemical reaction mechanism from batch process data. In Proceedings of the 17th IASTED International Conference on Modelling and Simulation (IASTED MS 2006), Paper 530-102, Montreal, Quebec, May 2006.
Presented Monday 17, 12:11 to 12:30, in session Process Simulation & Optimization - I (T4-9a).
