In this paper we present a new method for evaluating molecular similarity b
etween small organic compounds. Instead of a linear representation like fin
gerprints, a more complex description, a feature tree, is calculated for a
molecule. A feature tree represents hydrophobic fragments and functional gr
oups of the molecule and the way these groups are linked together. Each nod
e in the tree is labeled with a set of features representing chemical prope
rties of the part of the molecule corresponding to the node. The comparison
of feature trees is based on matching subtrees of two feature trees onto e
ach other. Two algorithms for tackling the matching problem are described t
hroughout this paper. On a dataset of about 1000 molecules, we demonstrate
the ability of our approach to identify molecules belonging to the same cla
ss of inhibitors. With a second dataset of 58 molecules with known binding
modes taken from the Brookhaven Protein Data Bank, we show that the matchin
gs produced by our algorithms are compatible with the relative orientation
of the molecules in the active site in 61% of the test cases. The average c
omputation time for a pair comparison is about 50 ms on a current workstati
on.