We present results of comparative modeling on 11 targets from the CASP3 exp
eriment. Our methods comprise the following steps: first, PSI-BLAST is used
to find homologues of the target sequence in the nonredundant GenBank prot
ein sequence database; second, after several iterations of PSI-BLAST, the r
esulting profile or position-specific similarity matrix is used to search a
database of Protein Databank (PDB) sequences; third, from the list of hits
resulting from the PDB search, a parent structure is chosen on the basis o
f the quality of the alignment and the quality of the experimental structur
e; fourth, this alignment is adjusted manually whenever insertions or delet
ions take place in secondary structure regions of the parent; fifth, the ba
ckbone is modeled from the parent structure and the alignment; and finally,
the program SCWRL is used to replace nonconserved side chains onto the par
ent backbone given the target sequence. For comparison, we also produced st
ructural models from the unaltered PSI-BLAST alignment, from an alignment f
rom the nonprofile version of BLAST, and from the global sequence alignment
program CLUSTAL W. Our results indicate that PSI-BLAST produced considerab
ly better alignments than would be possible with either global or local pai
rwise sequence alignment algorithms and that manual adjustments were helpfu
l. SCWRL, which uses a backbone-dependent rotamer library to predict side-c
hain conformations, did well in comparison with other methods used in CASP3
. (C) 1999 Wiley-Liss, Inc.