Learning Qualitative Models of the Detoxification Pathway of Methylglyoxal Wei Pang and George M. Coghill Department of Computing Science University of Aberdeen Aberdeen, AB24 3UE UK {pang.wei, g.coghill}@abdn.ac.uk Abstract In this paper, a qualitative model learning (QML) system is proposed to reconstruct the detoxification pathway of Methylgly- oxal. First an algorithm is implemented to convert possible pathways to qualita- tive models. Then a general learning strategy is presented. Finally a clonal selection algorithm is employed to adapt to future possible more complicated path- ways and its performance is compared with those of backtracking algorithms. 1 Introduction 1.1 Qualitative Model Learning Qualitative Model Learning (QML) is a branch of Qualitative Reasoning [6]. It involves ex- tracting qualitative models of dynamic systems from available observed data, which can be ei- ther quantitative or qualitative. The models that QML aims to infer are in the form of Qualitative Differential Equations (QDE). A QDE is an abstraction of a class of Ordinary Differential Equations (ODE), in the sense that a corresponding ODE can be obtained by parameterizing and quantizing this QDE. A QDE is the conjunctions of a set of qual- itative constraints, which link the variables in the model and express the relations among these variables. Qualitative values can be defined as either landmark values and intervals between these landmark values [7], or fuzzy intervals [14]. Some qualitative constraints are derived from mathematical relations, such as addition, sub- traction and multiplication ; others represent the incomplete knowledge about the function rela- tions under study, such as M + (monotonically increasing function) and M - (monotonically de- creasing function) in QSIM [7], which state that one variable will monotonically increase (de- crease) with the increase (decrease) of another. QML is particularly useful in the situation when little knowledge about the system is known and only sparse experimental data are available. In the last two decades, several QML systems have been developed, such as MISQ [12], QSI [13], ILP-QSI [2]. All these systems are general- purpose QML systems, but only ILP-QSI can in- tegrate domain-specific knowledge explicitly and easily. 1.2 Research on the Detoxification Pathway of Methylglyoxal Methylglyoxal (MG) is a naturally occurring toxic electrophile that is harmful to cells. Ex- cessive production of MG leads to cell death. Most of the organisms protect themselves from the toxic effect of MG through the detoxifi- cation pathway, which has been initially studied by Booth et al. [5]. The current understanding of the detoxifica- tion pathway is shown in Figure 1. When MG crosses the cell membrane and enters the cell, glutathione (GSH) reacts spontaneously with MG to produce hemithioacetal (HTA). Cat- alyzed by the enzyme glyoxalase I (GlxI), HTA is converted into S-lactoyl-glutathione (SLG). Cat- alyzed by a second enzyme glyoxalase I (GlxII), SLG is converted into GSH and non-toxic D- lactate. The whole pathway is composed of three biochemical reactions: the first is an non- enzymatic reaction which follows the law of mass action; the second and third are enzymatic reac- tions which obey Michaelis-Menten kinetics [9]. Based on the above understanding, a quanti- tative model [4] which consists of four differen- tial equations has been manually built. 1 How- ever, it should be pointed out that the mecha- nism of the MG detoxification is not yet fully understood, and the research is still ongoing. Consequently the quantitative model is an inter- 1. This quantitative model has not been published yet. So it will not be presented in this paper. Proceedings of the 2008 UK Workshop on Computational Intelligence Page 53