Development and comparison of stability and affinity based networks

FENOY, LUIS EMILIO; NIELSEN, MORTEN

Bahía Blanca

Congreso; VI Argentinean Conference on Computational Biology and Bioinformatics; 2015

A2B2C

BackgroundMHC class I molecules play a central role in the immune response mediated by cytotoxic T lymphocytes (CTLs). Their specificity has made them the central focus in the development of tools for identification of immunogenic peptides. Several in-silico methods predicting affinity binding of peptide-MHC-I have been developed in the last decades with excellent results[1]. However not every peptide that binds a MHC molecule is immunogenic. Since the MHC molecule not only has to bind the peptide, but also retain it until the arrival of the specific CTL clone that is able to recognize it, it has been speculated that the stability of peptide-MHC-I binding rather than the affinity its determinant of peptide immunogenicity[2,3,4]. In this work, we intend investigate this by training and comparing the predictive performance of pan-specific prediction tools based on stability and affinity data.Materials and methodsA dataset of 26910 binding half-life measures was used to train a pan-specific artificial neural network for stability predictions. The training was done as described in [5] combining peptide binding data with a pseudo sequence representation of the MHC class I protein sequence using a typical five-fold cross-validation scheme where four-fifths of the data were for training and the last fifth was for testing and early stopping. This was repeated five times so that all test sets were used for evaluation alternately. The resulting network was tested with epitopes and ligands from IEDB [http://www.iedb.org] and SYFPEITHI [http://www.syfpeithi.de] databases. Also, the predictions were compared and combined with a state-of-the-art affinity predictor: NetMHCpan[5]. In order to make a fair comparison of the performance of affinity-data and stability-data based networks, two new networks were trained, each one with the same amount of data-points and distribution of unique peptides. The performance was evaluated using the AUC value and combined affinity and stability predictions trough S=w*aff+(1-w)*stab, where S is the combined score, w a weight varying from 0 to 1, aff the prediction score of the affinity based network and stab the prediction score of the stability based network.ResultsWe were able to train a predictor with an excellent performance and the combination with NetMHCpan predictions significantly improved the performance for identification of MHC ligands. When trained balanced networks, the stability-based predictor showed a significant improvement prediction performance in comparison with the affinity-based predictor (Fig.1) reinforcing the idea that stability is a better immunogenicity predictor than affinity.