Biology Reference
In-Depth Information
MHCPEP and SYFPEITHI databases) and a set of 464 non-binders
(derived randomly from a human albumin sequence), demonstrated
that the BN shows up to a 99.0% accurate identification of the HLA-
A2 binding peptides. Peptides were encoded using 20-bin vector rep-
resentations, and HLA-A2 binding activities representations were the
same as those used by Brusic
et al
. (1994(b), 121-130). In addition,
their approach may also suggest certain advantages of probabilistic
graphical methods over other methods like the ANN. When the train-
ing data set was reduced by 40.0% from the original data, the predic-
tive accuracy (as measured by A
ROC
) of HLA-A2 binding peptides by
BN, HMM, and ANN methods was 0.88, 0.85 and 0.85, respectively.
Therefore, application of BN methods may provide an alternative
approach to immunologists when data are sparse. Astakhov and
Cherkasov clearly show that the BN method can potentially serve as
a machine-learning tool that can be applied to prediction of viral
immunogens as well as to
de novo
design strategies.
The studies described above demonstrate the power of ANN, as
well as other machine-learning methods, to exploit information from
peptide/protein databases in order to derive SAR and make predic-
tions of peptide/protein functions from primary sequences alone.
Most of the studies described herein used three-layered feed-forward
ANN with back-propagation methods. Expansion into other ANN
methods, such as self-organized maps,
272
self-growing neural net-
works,
273,274
and associative memory networks,
272
should bring
advances in peptidomimetics and
de novo
design. Also, development
of hybrid approaches that combine ANN methods with other
machine-learning techniques such as HMM, EA, BN, and SVM will
enhance prediction performance and advance development of vaccine
and diagnostic candidates. Neural network techniques are not just for
computer scientists or engineers, but comprise a powerful set of tools
for immunologists as well. Advancements in ANN applications to
immunology and protein design will depend on the ingenuity in for-
mulating problems and designing neural networks and the reliability
of the information stored in databases. We anticipate fast growth in
ANN applications for computational virology and in the number of
immunologists using such mathematical machine-learning techniques.
