The goal of protein domain structure prediction is to recognize
building blocks of the tertiary protein structure from knowledge of
the sequence alone. Knowledge on the location of these blocks on the
sequence can aid protein structure prediction and may be useful for
experimental purposes. A probabilistic model describing the connection
between a protein's domain structure and the output of multiple homology
searches is presented. Based on this model, the fully automatic protein
domain structure prediction server Dopro has been implemented. The
server was part of the CASP/CAFASP 2004 experiment, the results of which
will be discussed.