Information Technology Reference
In-Depth Information
∑
(
pIC
−
pIC
)
2
50
(exp)
50
(
est
)
(3)
2
r
=
1
−
i
=
1
∑
(
pIC
−
pIC
)
2
50
(exp)
50
(
mean
)
i
=
1
∑
=
(
pIC
−
pIC
)
2
50
(exp)
50
(
est
)
(4)
SEE
=
i
1
n
−
c
−
1
where
n
is the number of peptides and
c
is the number of components. In the present
case, a component in PLS is an independent trend relating measured biological activ-
ity to the underlying pattern of amino acids within a set of peptide sequences. In-
creasing the number of components improves the fit between target and explanatory
properties; the optimal number of components corresponds to the best
q
2
. Both SEP
and SEE are standard errors of prediction or estimation and assess the distribution
of errors between the observed and predicted (estimated) values in the regression
models.
4.2.4 Iterative Self-Consistent Algorithm - Class II Alleles
An ISC-PLS-based additive method was applied to the set of class II alleles. The ISC
PLS-based algorithm (Doytchinova and Flower 2003) works by generating a set of
nonameric subsequences extracted from the parent peptide. Values for pIC
50
corre-
sponding to this set of peptides were predicted using PLS and compared to the ex-
perimental pIC
50
value for each parent peptide. The best predicted nonamer was
selected for each peptide, i.e., those with the lowest residual between the experimen-
tal and predicted pIC
50
. LOO-CV was then employed to extract the optimal number
of components, which was then used to generate the non-cross-validated model.
Each new model is built from the chosen set of optimally scored nonamers. The
method works by comparing the new set of peptide sequences with the old set and if
the new set is different, the next iteration is begun. The process is repeated until the
set of extracted nonameric peptide sequences identified by the procedure have con-
verged. The resulting coefficients of the final non-cross-validated model describe the
quantitative contributions of each amino acid at each of the nine positions. An exam-
ple coefficient matrix for the I-A
b
allele is shown in Table 1.
4.2.5 Comparative Molecular Similarity Index Analysis (CoMSIA)
4.2.5.1 Molecular Modeling
Wherever possible an X-ray crystallographic structure for the nonameric/octameric
peptide binding to the various class I alleles was chosen as a starting conformation.
Using the crystallographic peptide as a template, all the studied peptides were built,
and then subjected to an initial geometry optimization using the Tripos molecular
force field and charges derived using the MOPAC AM1 Hamiltonian semiempirical
method (Dewar, Zoebisch, Healy, and Stewart 1985).
Molecular alignment was
based on the backbone atoms of the peptides, which was defined as an aggregate
during optimization.
