Combinatorial small molecule growth algorithm was used to design inhibitors for human carbonic anhydrase II. Two enantiomeric candidate molecules were predicted to bind with high potency (with R isomer binding stronger than S), but in two distinct conformations. The experiments verified that computational predictions concerning the binding affinities and the binding modes were correct for both isomers. The designed R isomer is the best-known inhibitor (K d ϳ 30 pM) of human carbonic anhydrase II. T he development of new drugs often depends on the identification of molecules (''leads'') that have high affinities for specified macromolecular targets. Although tight binding is only one important characteristic of a drug (1), it is often used as a guide in initial stages of drug discovery. Two contrasting methods-combinatorial (2-4) and rational (5-8)-represent the extremes in strategies for discovery of high-affinity leads. Combinatorial methods make it possible to screen large numbers of potential candidates and do not require prior knowledge of the structure of the receptor molecule. Rational methods attempt to design high-affinity ligands based on knowledge of the atomlevel structure of the receptor and of molecular interactions. In practice, both combinatorial and rational methods can efficiently identify relatively low-affinity leads; both are inefficient and unreliable in identifying high-affinity (K d ϳ nM) ligands (9-12).Here, we describe a computational methodology that combines combinatorial and rational strategies in the form of a computational system that is rapid enough to generate biased libraries of leads and accurate enough to give useful predictions of energetics and geometry. In its first experimental test, this method yielded a new ligand for a human carbonic anhydrase II (HCA) that has the highest known affinity for this enzyme (K d ϳ 30 pM). To our knowledge, it is the first time that a computational method has created a ligand that has the highest known affinity for a protein target.Our method, called CombiSMoG for combinatorial small molecule growth, incorporates the philosophy of combinatorial synthesis into computational drug design and is based on two interrelated components: a knowledge-based potential and a Monte Carlo ligand growth algorithm. The knowledge-based potential (13) is derived from a set of 1,000 protein-ligand complexes, whose structures are deposited in the Protein Data Bank crystallographic database (14). In this potential, two atoms-one on the ligand and one on the protein-are said to be in contact if the distance between them is less than a specified cutoff value (usually 5 Å). The contacts are classified according to the constituent atom types, and the frequencies of their occurrences in the database are transformed into energies by means of a Boltzmann-like relation to give the scoring function used in CombiSMoG (15,16). This potential has three main advantages over the commonly used force fields (17)(18)(19)(20). (i) The binary definition of atom-atom interactions all...