Pharmaceutical drug design is a long and expensive process. Early selection of promising molecules can dramatically improve this process, but this selection is often similar to searching for a needle in a haystack. In most cases, all that a chemist initially has is a small collection of molecules that exhibit enough desired activity to hypothesize that they share a 3D atomic substructure binding to the same receptor site. A key problem is to identify this substructure, which can then be used as a pattern to screen databases of molecules. This problem is complicated by the fact that a drug molecule is "flexible" and can achieve many low-energy (stable) states. We present search techniques that we have developed to find these low-energy states and to identify common 3D substructures that appear in at least one low-energy state of most molecules in a given collection. We also show experimental results obtained with a software system implementing these techniques.