Currently available sequence alignment programs are generally not capa
ble of detecting functional and structural homologs in the twilight zo
ne of sequence similarity, i.e. when the sequence identity falls below
about 25%. Here we attempt to detect such weak similarities using an
approach based on a notion of protein sequence similarity radically di
fferent from that used in sequential alignment. The approach defines p
rotein sequence dissimilarity (or distance) as a weighted sum of diffe
rences of compositional properties such as singlet and doublet amino a
cid composition, molecular weight, isoelectric point (protein property
search or PropSearch). With PropSearch, either single sequences can b
e used for a database query, or multiple sequences can be merged into
an ''average'' sequence reflecting the average composition of a protei
n family. First, we show that members of structural protein families h
ave a low mutual PropSearch distance when the weights are optimized to
discriminate maximally between structural families. Second, we demons
trate the results of database searches using the PropSearch method. Su
ch searches are very rapid when scanning a preprocessed database and d
o not require alignments. In cases in which conventional alignment too
ls fail to detect similarities, PropSearch can be used to generate hyp
otheses about possible structural or functional relationships between
a new sequence and sequences in the database.