In the search for new drugs, it often occurs that the binding affiniti
es of several compounds to a common receptor macromolecule are known e
xperimentally, but the structure of the receptor is not known. This ar
ticle describes an extraordinarily objective computer algorithm for de
ducing the important geometric and energetic features of the common bi
nding site, starting only from the chemical structures of the ligands
and their observed binding. The user does not have to propose a pharma
cophore, guess the bioactive conformations of the ligands, or suggest
ways to superimpose the active compounds. The method takes into accoun
t conformational flexibility of the ligands, stereospecific binding, d
iverse or unrelated chemical structures, inaccurate or qualitative bin
ding data, and the possibility that chemically similar ligands may or
may not bind to the receptor in similar orientations. The resulting mo
del can be viewed graphically and interpreted in terms of: one or more
binding regions of the receptor, each preferring to be occupied by va
rious sorts of chemical groups. The model always fits the given data c
ompletely and can predict the binding of any other ligand, regardless
of chemical structure. The method is an outgrowth of distance geometry
and Voronoi polyhedra site modeling but Incorporates several novel fe
atures. The geometry of the ligand molecules and the site is described
in terms of intervals of internal distances. Determining the site mod
el consists of reducing the uncertainty in the interregion distance in
tervals, and this uncertainty is described as intervals of intervals.
Similarly, the given binding affinities and their experimental uncerta
inties are treated as intervals in the affinity scale. The final site
model specifies an entire region of interaction energy parameters that
satisfy the training set rather than a single set of parameters. Pred
icted binding for test compounds results in an interval which, when co
mpared to the experimental interval, may be correct, incorrect, or vag
ue. There is a pervasive ternary logic involved in the assessment of p
redictions, in the search for a satisfactory model, and in judging whe
ther a given molecule may bind in a particular orientation: true, fals
e, or maybe. The approach is illustrated on an extremely simple artifi
cial example and on a real data set of cocaine analogues binding to a
nerve membrane receptor in vitro. (C) 1995 by John Wiley and Sons, Inc
.