Application of abductive ILP to learning metabolic network inhibition from temporal data

In this paper we use a logic-based representation and a combination of Abduction and Induction to model inhibition in metabolic networks. In general, the integration of abduction and induction is required when the following two conditions hold. Firstly, the given background knowledge is incomplete. Secondly, the problem must require the learning of general rules in the circumstance in which the hypothesis language is disjoint from the observation language. Both these conditions hold in the application considered in this paper. Inhibition is very important from the therapeutic point of view since many substances designed to be used as drugs can have an inhibitory effect on other enzymes. Any system able to predict the inhibitory effect of substances on the metabolic network would therefore be very useful in assessing the potential harmful side-effects of drugs. In modelling the phenomenon of inhibition in metabolic networks, background knowledge is used which describes the network topology and functional classes of inhibitors and enzymes. This background knowledge, which represents the present state of understanding, is incomplete. In order to overcome this incompleteness hypotheses are considered which consist of a mixture of specific inhibitions of enzymes (ground facts) together with general (non-ground) rules which predict classes of enzymes likely to be inhibited by the toxin. The foreground examples are derived from in vivo experiments involving NMR analysis of time-varying metabolite concentrations in rat urine following injections of toxins. The model's performance is evaluated on training and test sets randomly generated from a real metabolic network. It is shown that even in the case where the hypotheses are restricted to be ground, the predictive accuracy increases with the number of training examples and in all cases exceeds the default (majority class). Experimental results also suggest that when sufficient training data is provided, non-ground hypotheses show a better predictive accuracy than ground hypotheses. The model is also evaluated in terms of the biological insight that it provides.