Prediction of possible beginning-intermediate-terminating metabolite transformation pattern using supervised neural networks and chemical properties

Abstract

The problem of atomically transformable metabolite route by using the supervised learning paradigm was studied. This problem in forms of specifying whether the considered beginning−(intermediate)−terminal metabolite queries are related to any transferable path and was treated as supervised classification problem. A defined metabolite query set and a class target set were obtained from graph path searching strategy in a predefined biochemical transforming graph of reactions and their associated metabolites. For each metabolite query, input data were the transformed features earned by using a computed molecular property set of all metabolites in that query. A method to treat the unequal proportions of binary class training data in all four defined problems was developed. Consequently, the proposed method was tested with the E.coli compound transformation networks including comparative tasks with the unseen data. The four predictive models yielded acceptable performance results i.e. G-mean ~ 0.77−0.93 at output cut-off value = 0.4. In comparative works, the results concluded that both the improved original values and the mean of the improved mean output values provided by adding information about existing/non-existing compound structure alignment showed the generally better performance in all four defined questions.