Large-scale computational identification of HIV T-cell epitopes

Bioinformatics-driven T-cell epitope-identification methods can enhance vaccine target selection significantly. We evaluated three unrelated computational methods to screen Pol, Gag and Env sequences extracted from the Los Alamos HIV database for HLA-A*0201 and HLA-B*3501 T-cell epitope candidates. The hidden Markov model predicted 389 HLA-B*3501-restricted candidates from 374 HIV-1 and 97 HIV-2 sequences. The artificial neural network (ANN) model, and Bioinformatics and Molecular Analysis Section (BIMAS) quantitative matrix predictions for A*0201 yielded 1122 HIV-1 and 548 HIV-2 candidates. The overall sequence coverage of the predicted A*0201 T-cell epitopes was 2.7% (HIV-1) and 3.0% (HIV-2). HLA-B*3501-predicted epitopes covered 0.9% (HIV-1) and 1.4% (HIV-2) of the total sequence. Comparison of 890 ANN- and 397 BIMAS-derived HIV-1 A*0201-restricted epitope candidates showed that only 13-19% of the predicted and 26% of the experimentally confirmed T-cell epitopes were captured by both methods. Extrapolating these results, we estimated that at least 247 predicted HIV-1 epitopes are yet to be discovered as active A*0201-restricted T-cell epitopes. Adequate comparison and combined usage of various predictive bioinformatics methods, rather than uncritical use of any single prediction method, will enable cost-effective and efficient T-cell epitope screening.