Table 5.

LOO benchmark comparison of the pan-specific NetMHCIIpan-2.0 and the NetMHCIIpan-1.0 methods

			NetMHCIIpan-1.0		NetMHCIIpan-2.0		TEPITOPE
Allele	#	#bind	PCC	AUC	PCC	AUC	AUC
DRB1*0101	5166	3510	0.571	0.778	0.627	0.794	0.720
DRB1*0301	1020	277	0.465	0.746	0.560	0.792	0.664
DRB1*0401	1024	510	0.591	0.775	0.652	0.802	0.716
DRB1*0404	663	386	0.693	0.852	0.731	0.869	0.770
DRB1*0405	630	425	0.594	0.808	0.626	0.823	0.759
DRB1*0701	853	498	0.655	0.825	0.753	0.886	0.761
DRB1*0802	420	148	0.637	0.841	0.700	0.869	0.766
DRB1*0901	530	254	0.406	0.653	0.474	0.684
DRB1*1101	950	429	0.580	0.799	0.721	0.875	0.721
DRB1*1302	498	199	0.323	0.658	0.337	0.648	0.652
DRB1*1501	934	450	0.533	0.738	0.598	0.769	0.686
DRB3*0101	549	75	0.449	0.716	0.474	0.733
DRB4*0101	446	200	0.448	0.724	0.515	0.762
DRB5*0101	924	478	0.627	0.831	0.722	0.879	0.686
Ave			0.541	0.768	0.606	0.799
Ave*			0.570	0.786	0.639	0.819	0.718

The two methods are compared in a leave-one-out experiment on the peptide binding data described in the original NetMHCIIpan publication [29].

# is the number of peptide binding data for each allele, #bind is the number of peptides with a binding affinity stronger than 500 nM. NetMHCIIpan-1.0 is the method by Nielsen et al. [29], NetMHCIIpan-2.0 is the method described here, and TEPITOPE is the method by Sturniolo et al. [1]. Prediction values for NetMHCIIpan-1.0 were taken from [29]. Ave gives the per allele average, Ave* gives the per allele average of the 11 alleles characterized by the TEPITOPE method. In bold is highlighted the best performing method for each of the 14 alleles. AUC values were calculated using a binding threshold of 500 nM. Only AUC values are included for the TEPITOPE method since prediction values for this method are not linearly related to the binding affinity.