Biology Reference
In-Depth Information
The ANN had a predictive accuracy of 78% for binding to HLA-A2
and 88% for binding to mouse H-2K
b
.
Milik and colleagues (1998) applied the ANN method to the pre-
diction of mouse class I H-2K
b
using 8-mer peptides from a phage
display library.
205-207
Binding properties were measured by a compet-
itive CTL assay and values expressed as the amount of competitor
peptide (OVA 257) necessary to induce 50% inhibition.
208
They
obtained sequences of 181 binding and 129 non-binding phage. For
the classification procedure, the data were divided into three sub-sets:
training set consisting of 93 binding and 130 non-binding peptides;
a validation set 1 (18 binding and 13 non-binding), and validation set
2 (33 binding and 23 non-binding). The validation set 1 was used for
fine-tuning the system during the learning phase, and validation set 2
was used for evaluating the performance of the three-layer feed-for-
ward back-propagation ANN. They tested two ANN models based on
two different representation schemes for encoding amino acids. The
first scheme (ANN1) consisted of a 10-number representation based
on the hierarchical organization of amino acids
209
; the second
(ANN2) consisted of a six-number representation based on normal-
ized physicochemical properties of amino acids using volume, bulki-
ness, flexibility, polarity, aromaticity, and charge. ANN1 consisted of
one input layer with 180 nodes, one hidden layer with five hidden
nodes, and one single output neuron, whereas, ANN2 consisted of 48
input neurons, 3 hidden neurons, and 1 output neuron. To evaluate
the validity of the analysis and trained ANN, they predicted H-2K
b
binding peptides from the sequence of chicken albumin, a protein
with a well-characterized H-2K
b
epitope. All peptides that were clas-
sified as “binding” were synthesized and tested for experimentally.
ANN1 and ANN2 selected the majority of binding peptides tested,
with sensitivities of 70% and 80%, respectively.
Honeyman
et al
. (1998) and Brusic
et al
. (1998b; 121-130)
reported high accuracy in the prediction of binding peptides (9-mers)
to the class II HLA-DR4(B1*0401). Their approach was different
from those previously discussed for the class I alleles. In both cases,
a hybrid method was developed for predicting binding peptides to MHC
class II molecules. Their method, termed PERUN, was implemented
