A distance geometry based protein modelling algorithm is presented whi
ch relies on the projection of simple model chain coordinates into Euc
lidean spaces with gradually decreasing dimensionality. Fast embedding
was achieved by performing separate distance matrix projections on su
bsets of the model points. Structural equivalences between the unknown
target and related proteins with known structures were deduced either
from a mixed sequence/structure multiple alignment or from the output
of various fold recognition (threading) approaches. These equivalence
s were mapped onto the model as structure-specific conserved C-alpha a
tom distances and secondary structure assignments. Additional nonspeci
fic distance restraints derived from general stereochemical properties
of folded protein chains were used to guide the modelling process. Th
e method quickly constructed a large number of low-resolution models w
hich could then serve as starting conformations for full-atom refineme
nt. Structure predictions for some targets in the 'Asilomar Challenge'
(GASPS) are presented to illustrate potential applications of the app
roach.