Chem. Pharm. Bull. 55(7) 980—984 (2007)

into pharmaceutical research in recent years to provide information about potential drug targets and their disease association. In most drug discovery projects, however, we are still in the situation where one does not have any specific target in mind at the time when a set of molecules for screening is necessary. In this situation, we often use small number of active compounds as clues to discover better drug candidates. Since quite a large number of chemicals are commercially available, the discovery process usually starts from screening drug candidates in these chemical databases. In pharmaceutical companies, the proprietary compounds are also added to the target databases to be screened. A tremendous number of compounds is usually contained in such databases. Therefore, it is usually not realistic that even modern highthroughput screening technique can actually handle all of the compounds. If it is possible to reduce the number of compounds to be experimentally tested, it can make the entire process of searching cost effective and fast. Virtual screening can serve as a tool to suggest compounds to be tested experimentally. In this sense, virtual screening is a computational counterpart to high-throughput screening. As virtual screening is an entirely computer-based method, it can search a very large chemical space. Virtual screening is becoming an important screening technique because it is also very cost effective. In this paper, a simple method of virtual screening is proposed. It is usually difficult to identify the specific conformation of a drug that is responsible to the biological activity. Accordingly, in many drug discovery projects, we have no alternate but to use only two dimensional chemical structures. Since chemical characters of drug molecules determine their drug-likeness, this method uses only chemical characters calculated from the two-dimensional chemical structures of the molecules. Local and global chemical characters are represented by molecular fingerprint and trait, respectively. The concept of molecular fingerprint is relatively well-established in the field of cheminformatics. As a molecular fingerprint is expressed by substructural information, the molecular fingerprint represents a local chemical character of a molecule. The trait, however, is newly introduced concept in this study and it can represent a global character of a set of molecules. A trait of a particular group of molecules is expressed by a combination of various two dimensional (2D) molecular descriptors. The 2D molecular descriptors only use the atoms and connection information of the molecule. The trait represents common global characters of these molecules. A druglikeness profile of a set of drugs with a particular pharmacological activity can be expressed by a combination of the molecular fingerprint and the trait. To learn about drug-likeness of drugs with a specific pharmacological activity, we used a database of drugs that are clinically used in Japan (DCUJ) now. The number of independent drugs in DCUJ was 1170 when we undertook this study. Namely, DCUJ was used as a training dataset. The drug-likeness profile trained on these real drugs was employed to predict the specific drug-likeness of compounds in other chemical databases. In these chemical databases which consist of thousands of compounds, an appreciable number of compounds that show the relevant pharmacological activity are contained. Some of these compounds are drugs clinically used abroad, but not in Japan. The prediction rate was judged by an enrichment factor. Despite the simplicity of the methodology, reasonably practical results have been successfully obtained.