In Silico QSAR-Based Predictions of Class I and Class II MHC Epitopes - Immunoinformatics - page 68

hydrophilic interactions enhance affinity). The hydrogen-bond donor map is shown
in cyan (donors on the ligand are preferred) and purple (donors are disfavored)
contours. Finally, in the hydrogen-bond acceptor map favored areas are in magenta
and disfavored in yellow.

4.3 Results

The Additive, ISC-PLS, and CoMSIA models were generated for 23 human and
mouse class I and class II alleles.

4.3.1 Additive Method - Class I Alleles

The generated models ( n =30-335), as shown in Table 2, have an acceptable level of
predictive power: LOO-CV statistical terms, SEP and q 2 , ranged between 0.565 and
0.907, and 0.317 and 0.531 respectively. The non-cross-validated statistical terms
NC, SEE, and r 2 ranged between 2 and 9, 0.085 and 0.456, and 0.731 and 0.997,
respectively. An extended motif, as defined by the class I models, is summarized in
Table 3. It shows anchor and nonanchor residues relating to strong and weak binding
residues. For simplicity, the quantitative contributions of amino acids at each posi-
tion for the class I mouse alleles are shown in Fig. 1.

Table 2. Class I Additive-PLS Method results.

LOO Non-cross Validation

n a

Epitope

SEP b q 2 c NC d SEE e r 2

A*0101 95 0.907 0.420 4 0.146 0.997

A*0201 335 0.694 0.377 6 0.456 0.731

A*0202 69 0.606 0.317 9 0.193 0.943

A*0203 62 0.841 0.327 6 0.197 0.963

A*0206 57 0.576 0.475 6 0.085 0.989

Human

A*0301 72 0.680 0.436 6 0.181 0.959

A*1101 62 0.572 0.458 2 0.321 0.829

A*3101 30 0.710 0.482 3 0.325 0.892

A*6801 38 0.594 0.531 4 0.175 0.959

A*6802 46 0.647 0.500 7 0.119 0.983

B*3501 52 0.710 0.435 6 0.118 0.984

H2-K k

154 0.565 0.456 6 0.198 0.933

H2-K b

Mouse 62 0.894 0.454 6 0.128 0.989

H2-D b 65 0.837 0.493 5 0.268 0.948
a Number of epitopes; b Standard Error of Prediction; c from Leave-One-Out Cross-Validation;
d Number of components; e Standard Error of Estimate.