Highlights: Predicting the cross-immunoreactivity of hepatitis C virus hyper-variable region 1 peptides using polynomial neural networks

Hepatitis C is a liver disease caused by infection with hepatitis C virus (HCV). The number of people infected with HCV is estimated at ∼130 million word-wide. Approximately, 70%–80% of HCV-infected patients fail to clear the virus and develop chronic hepatitis C (CHC), which is a well-known risk factor for development of liver fibrosis, cirrhosis and hepatocellular carcinoma (HCC). At the present time, there are no preventive vaccines against HCV. In spite of the different strategies used for vaccine development against HCV, the development of a successful vaccine remains elusive. The HCV genome exhibits high degree of genetic diversity and divergence, which, in addition to HCV's rapid rate of mutation and capacity to escape from the host immune response, has hindered efforts towards the formulation of a broadly protective vaccine. For the study of humoral immunity responses to HCV infection, synthetic peptides derived from the HVR1 of the envelope protein 2 (E2) of HCV genome have been used as immunogens in small animals and non-human primates. HVR 1 peptides contain a neutralizing epitope and are typically very immunogenic in mice and non-human primates although the pattern of reactivity varies significantly depending on the HCV strain. The challenge of developing an efficacious vaccine(s) is finding immunogens capable of eliciting a humoral response that leads the immune system of the host to produce antibodies (Ab's) which broadly and strongly cross-immunoreact against all HCV strains circulating in human populations. The immune-reactivity properties of HCV proteins can be modeled in quantitative-structure-activity-relationship (QSAR) models. Polynomial neural networks, as non-parametric algorithms offer an ideal approach to QSAR studies. In this study, we built polynomial neural networks (PNN) to model the immune-reactivity properties of synthetic HCV HVR1 peptides and examined PNN performance to predict cross-immunoreactive patterns using amino-acid (aa)-based physicochemical parameters. We implemented a non-linear regression PNN algorithm to correlate aa-based physicochemical properties of HVR1 peptides to their corresponding cross-immunoreactivity patterns. PNN models were synthesized from a dataset of hyper-immune sera tested against 326 HVR1 peptide variants (representative of 6 genotypes of HCV). Synthesis of PNN models was constrained to 20 iterations and degree= x*y. The residual sum of squares (RSS) criterion was used for selection of the best PNN model. Evaluation of PNN models was conducted using the reliability rate (RR) criterion on two external validation datasets (n=4 and n=10). The statistics of training evaluation of the best PNN models was: R^2=0.771, q^2=0.771, RMSE=1.7776, var=0.229 and MAE=1.305, where R^2 is the square of correlation coefficient; q^2, the squared correlation coefficient of predictions; var, the variation accuracy criterion; RMSE, the Root Mean Squared Error as and MAE, the Mean Absolute Error. The statistics of performance evaluations on validation set (n=4) was: R^2=0.937, q^2= 0.748, RMSE=1.8316, var=0.252, MAE=1.294, and for validation set (n=10): R^2=0.632, q^2=0.565, RMSE=2.7244, var=0.435, MAE=2.059. Furthermore, linear projection (LP) mappings indicate differential spatial distribution between cross-reactive and non-cross-reactive HVR1 peptides with high sensitivity and specificity (classification accuracy = 0.9967, sensitivity = 1.0; specificity = 0.9383; F score = 0.9983; Brier score = 0.0036). In addition, several HVR1 features relevant for accurate performance of models were identified. Our findings indicate that the immune-reactivity property of the HCV HVR1 is associated to the physicochemical profile of peptide sequences. Such association can be modeled by mathematical and computational means. Predictive models generated in this study may provide useful aid for selection of potentially efficacious HCV HVR1 vaccine candidates.