Biology Reference
In-Depth Information
been reported to outperform binding motif and QM-based meth-
ods.
161,170,203
Predictive performances using ANN methods for the
prediction of MHC-binding peptides are in the order of 80% in sen-
sitivity and specificity.
161,163,204
Prediction of MHC class I and class II binders with ANN
.
Brusic and colleagues (1995) were the first to report on ANN meth-
ods for prediction of MHC-binding peptides.
119
They trained a fully
connected three-layer back-propagation ANN to classify peptides into
those predicted to bind and not to bind to either human HLA-A2 or
mouse H-2K
b
. The number of input nodes was equal to the length of
the input signal (nine nodes for 9-mers or eight nodes for 8-mers rep-
resenting HLA-A2 and H-2K
b
, respectively); the number of hidden
neurons was nine, and one output node predicted binding versus
non-binding. Training samples were obtained from the MHCPEP
database. The training set consisted of 186 positive examples (bind-
ing peptides) and 1,071 negative examples (non-binding peptides) for
the HLA-A2 predictions; and 30 binders and 796 non-binders for the
H-2K
b
predictions. For the encoding of peptide vectors, three signal
input representations (Table 1) were used: a 20-bin representation in
which each amino acid was encoded by a unique binary string of 0
and 1; Rep6, a six-number representation based on a scalar value for
a physicochemical feature; and Rep9, a nine-number intermediate
representation using a feature-based grouping of amino acids. During
ANN training, each signal input in the training set was mapped to
corresponding binding affinities, represented by IC
50
values, which
were defined as the concentration of peptide required to inhibit the
binding of a universal DR4-binding peptide by 50%. The system was
validated by testing the ANN performance with selected 9-mers from
published data obtained by
in vitro
assays.
195
The evaluation set con-
sisted of 60 known binders and 68 non-binders for HLA-A2 and 23
binders and 77 non-binders for H-2K
b
. Of the 128 peptides in the
evaluation set for HLA-A2 predictions, 30 were derived from natural
proteins and the remainder from synthetic peptides, including poly-
glycine and poly-alanine sequences. The H-2K
b
predictions used two
test groups: one with nine known binders from different proteins, and
the other with 14 binders and 77 non-binders from natural proteins.
