Comparative analysis of binary logistic regression to artificial neural networks in predicting precursor sequence cleavage

Abstract

Bioinformatic predictions of neuropeptides resulting from enzymatic cleavages of precursors enable a range of follow-up studies that are aided by accurate predictions. A comparative study of the performance of complementary cleavage prediction models has been undertaken. Binary logistic and artificial neural network (ANN) models were created using various strategies and trained and tested on bovine and rat precursors with experimental cleavage information. Multiple criteria were used to compare 4 logistic regression models with varying properties and 8 ANN with varying structures. All models had high specificity (>90%) and sensitivity ranged from 68% to 100%. ANN based on well-represented amino acid locations performed similarly or slightly worse than networks based on all amino acid locations. Logistic parameter estimates aided in the identification of amino acids associated with cleavage. No model was superior across data sets and thus, prediction of neuropeptides should rely on multiple model specifications and comprehensive training data sets.