Rational Design of Viral Protein Structures with Predetermined Immunological Properties - Structure-Based Study of Viral Replication - page 225

databases such as MHCPEP, 165 JenPep, 166 IMGT/HLA, 158,159,167
SYFPEITHI 168 and MHCBN. 169 Such a vast amount of information
serves as the basis for empirical computational approaches to elucidate
the physicochemical and structural parameters that regulate peptide
binding to MHC molecules.
History of T-cell epitope prediction. Algorithms for prediction
of T-cell epitopes can be classified as direct or indirect methods. 170
Direct methods are based on information about the T-cell epitopes,
whereas indirect methods are based on information about the MHC-
binding peptides. In the past, direct prediction methods relied on the
identification of amphipathic structures. 171-174 The AMPHI algorithm
is based on this assumption. 175 Another algorithm, SOHHA, was
developed based on the assumption that T-cell epitopes consist of
three to five helical turns with a narrow strip of hydrophobic residues
on one side. 176 Algorithms based on secondary structure analysis were
superseded by the identification of common motifs among T-cell epi-
topes in primary structures. Rothbard and Taylor (1988) collected
about 57 T-cell epitopes and, based on their patterns, published a list
of motifs. 177 They developed an algorithm based on the association
between cysteine-containing T-cell epitopes and certain other
residues. The system searches for triplets, including CAK, CLV, CKL
and CGS, in the peptide sequence. 178 In 1995, two computational
T-cell epitope prediction tools, EpiMer and OptiMer, were developed
based on knowledge of MHC binding motifs. 179 OptiMer predicts
amphipathic segments of protein with high motif density, and EpiMer
locates segments of protein with high motif density. However, when
tested against databases of human and murine T-cell epitopes, these
direct prediction methods performed with low accuracy. 180 In the last
decade, several indirect methods have been developed that predict
MHC binders instead of T-cell epitopes. 181-184 Due to more specific
interactions of MHC molecules and peptides, these methods are more
accurate than direct T-cell epitope prediction methods. 170 These indi-
rect methods are based on: 1) structural comparative 185 and Ab initio 186
modeling; 2) binding motifs 162,181-184,187-190 ; 3) quantitative matrices
(QM) 191-199 ; and more recently, 4) machine learning techniques like
neural networks. 163,170,200-202 The application of ANN algorithms has