Skip to Main Content
This paper describes a neural network based inference system developed as part of a bioinformatic application in order to help implement a systematic search scheme for the identification of genes which encode enzymes of metabolic pathways. The inference system uses BLAST sequence alignment values as inputs and generates a classification of the best candidates for inclusion in a metabolic pathway map. The system considers a workflow that allows the user to provide feedback with their final classification decisions. These are stored in conjunction with analyzed sequences for re-training and constant inference system improvement. The construction of the inference system involved the study of various neural topologies and training data models. Of the many training data models analyzed three are currently presented for comparison: using the BLAST algorithm's parameters directly, using standardized parameters determined by human experts, and a new proposal for input parameter normalization. The neural network was tested with all three data models. The three models enabled the inference system to perform a satisfactory rating of the candidates. Our proposal for parameter normalization produced the best model with several data sets showing a highly accurate prediction capability.