Use of a neural network to assign serologic specificities to HLA‐A, ‐B and ‐DRB1 allelic products

Publisher Website
Google Scholar
Abstract
A computational method was used to predict the serologic specificities of HLA molecules encoded by the HLA‐A, ‐B, and ‐DRB1 loci. The polypeptide sequences of a subset of alleles (numbering 149) with well‐defined serologic assignments were used to train a neural network to predict broad and split serologic assignments for each HLA allelic product. The resultant neural network assignments were compared with those of a validation set containing the sequences of 74 HLA‐A, 175 HLA‐B, and 117 HLA‐DRB1 alleles that had previous serologic test assignments but were not part of the training set. The network was able to correctly predict at least one of the serologic assignments of the majority of the validation alleles (99% of the HLA‐A set, 86% HLA‐B, 94% HLA‐DRB1). The remainder received either no assignment (1% HLA‐A, 13% HLA‐B, 5% HLA‐DRB1) or a different but closely related assignment (1% HLA‐B and ‐DRB1). Overall, the variation in serologic assignment by the network appeared comparable to the assignments seen among different laboratories using serologic techniques. When used to predict the serologic assignments of 393 HLA alleles without known serologic types, the network was able to predict assignments for most alleles (95% HLA‐A, 85% HLA‐B, 96% HLA‐DRB1). The majority of these assignments were consistent with assignments predicted by sequence homologies with known alleles. The remainder did not receive an assignment and likely represent new combinations of epitopes.