Biology Reference
In-Depth Information
peptides was 77% sensitivity and 80% specificity. Meanwhile, Brusic
et al
. (1998b, 121-130) compared the PERUN method with the QM-
based and binding motif-based methods for prediction of low, mod-
erate and high affinity binders. PERUN slightly outperformed the
QM method, with prediction performance of 73% for low binders,
86% for moderate binders and 88% for high affinity binders, whereas
the QM performance was 73%, 82% and 87%, respectively. The bind-
ing motif method had the lowest performance (63%, 69% and 74%,
respectively). Such an approach integrates the strengths of the differ-
ent methods and minimizes their disadvantages. For instance, binding
motifs encode the most important rules of peptide-MHC binding
interaction.
184,189,196,211
However, for class II molecules, binding motifs
are less well defined.
189
MHC class II molecules show degenerate
motifs, which makes peptide alignment more difficult. Because of
this, except for certain molecules,
196,211
binding motif-based methods
show poor generalizations. In contrast, QM-based methods, which
are in essence refined binding motifs, can predict large subsets of
binding peptides reasonably well and can be used when data sets are
limited.
192,193,195,196,212,213
However, QM cannot deal with nonlinearity
in data and may miss distinct subsets of binding preferences, such as
medium binders.
200
Also, QM methods are not adaptive and self-
learning, so integration of new data usually requires redesigning of
the alignment matrix. On the other hand, ANNs can deal with non-
linearity and are adaptive and self-learning, but usually require a large
amount of data.
A limitation of ANN predictions based on the indirect methods
described above is their inability to discriminate between T-cell epi-
topes and non-epitope MHC binders, mainly because these methods
only predict the MHC binders from antigenic sequences. Bhasin and
Raghava (2004) recently developed a direct method for predicting
cytotoxic T lymphocyte (CTL) epitopes from antigenic sequences
using a hybrid method that combines QM, support vector machine
(SVM), and ANN methods. Their system is based on 1,137 experimen-
tally proven CTL epitopes and 1,134 non-epitopes (786 non-epitopes
9-mers and 348 MHC non-binders). The system was evaluated with
a blind data set consisting of 63 CTL epitopes, 63 non-epitope
