Artificial neural network weights of residues for the serological specificities of HLA

Abstract

P>New human leucocyte antigen (HLA) alleles are assigned largely based on their sequence homologies due to lack of information on the serological reactivities. Artificial neural networks (ANNs) have been tested as a possible tool for helping to predict the serology of new alleles in the absence of serological information. However, an ANN analysis per se imparts no information regarding which residues are important in determining serological specificity. To address this issue, we extracted ANN weights of HLA residues. The ANN was trained using 139 HLA-A, 302 HLA-B and 136 HLA-DRB1 alleles, for which serological specificities were assigned in the 2004 Nomenclature Report. When the trained ANN was evaluated using alleles that were contained in the HLA Dictionary 2008 but had not been employed in the training set, the accuracy was 91% (29/32) for HLA-A, 91% (40/44) for HLA-B and 90% (9/10) for HLA-DR. Finally, ANN residue weights were extracted by summing the weights of directly connected ANN nodes. When we assessed the significance of the ANN residue weights by comparing our data with the results of epitope studies conducted by El-Awar and colleagues, we found that the ANN weights tended to be high at the epitopes described by El-Awar et al. Furthermore, the ANN weights extracted in this work could be used to explain ambiguous characteristics of serological specificities to some extent. Our data are thus considered to support the results of the epitope studies conducted by El-Awar and advance our understanding of the ambiguous serological specificities of some alleles.