Table 4.

Comparison of A_ROC values on HLA-DRB1*0401 data sets from MHCBench

Method	Set1	Set2	Set3a	Set3b	Set4a	Set4b	Set5a	Set5b	Avg.
TEPITOPE	0.776	0.740	0.740	0.754	0.763	0.750	0.651	0.661	0.729
PERUN	0.771	0.685	0.693	0.713	0.724	0.672	0.695	0.714	0.708
Gibbs Sampler2	0.803	0.775	0.75	0.762	0.793	0.787	0.6211	0.6611	0.744
LP_top22	0.725	0.721	0.728	0.753	0.719	0.728	0.8151	0.8591	0.756
SKM	0.870	0.832	0.823	0.821	0.862	0.827	0.787	0.770	0.824

Comparing performance of SKM with results reported for the Gibbs Sampling method [27], "LP_top2" [21], and PERUN [7]. Best results shown in bold.

1: Best reported results, where Cysteines are treated as Alanines [27].

2: Best reported results of [21].

Results of the LP_top2 and Gibbs Sampler are from evaluation on the MHCBench sets. However, as is described in [21], training was performed on a training set consisting of selected samples from MHCPEP [1] and SYFPEITHI [53]. However, MHCBench mainly consists of samples from MHCPEP, and a large overlap exist between training and test sets (e.g. 502 of 646 samples of Set 4a).