Comparison of prohormone processing prediction using different artificial intelligence algorithms

Neuropeptides are critical messaging molecules that are involved in numerous biological processes underlying diseases and behavior. Prediction of the final bioactive neuropeptides from a prohormone prone to errors because synthesis of neuropeptide involves cleavage and other post-translational modifications of large prohormone proteins. Reliable neuropeptide identification is restricted by the limited number of prohormone genes and experimentally reported cleavage sites. To address these limitations, different artificial intelligence (AI) algorithms and datasets were systematically evaluated using a test-training-validate stratified k-fold cross-validation strategy. Datasets were generated from human data and augmented with protein isoform sequences and sequences from primate species. Inclusion of protein isoforms improved model metrics including sensitivity, specificity, and P4, compared to using a single protein sequence per human prohormone gene. Model metrics were also improved for most AI algorithms using a hominoid dataset (11 hominoid species). Incremental improvements were observed in the model metrics when using primate dataset (63 hominoid, cercopithecoid, and platyrrhine species) over the hominoid dataset. The phenomenon of quasi-complete separation, characterized using logistic regression approach, was present in all algorithms that restricted the generalization of models trained on smaller datasets. Within the test datasets, all algorithms performed similarly for the human datasets. With larger hominoid and primate datasets, random forest, neural networks and 1-D convolutional neural networks had superior model metrics to other algorithms and datasets. This highlights the importance of dataset size in model training. The developed strategy provides the foundation for improved cleavage prediction and is applicable to enhance other genomic prediction problems.