Fragment reweighting in ligand-based virtual screening

• Main Author: Alfakiabdalla Abdelrahim, Ali Ahmed
• Format: Thesis
• Language: English
• Published: 2013
• Subjects: QD Chemistry
• Online Access: http://eprints.utm.my/id/eprint/33788/5/AliAhmedAlFakiabdallaPFSKSM.pdf; http://eprints.utm.my/id/eprint/33788/; http://dms.library.utm.my:8080/vital/access/manager/Repository/vital:69923?site_name=Restricted Repository
• Institution: Universiti Teknologi Malaysia

Based on the molecular similarity principle, functionally similar molecules are sought by searching molecular databases for structurally similar molecules to be used in rational drug design. The conventional 2-dimentional similarity methods are the most used methods to measure similarity of molecules, including fragments that are not related to the biological activity of a molecule. The most common methods among the 2-dimentional similarity methods are the vector space model and the Bayesian networks, which are based on mutual independence between fragments. However, these methods do not consider the importance of fragments. In this thesis, four reweighting approaches are proposed to identify the important fragments. The first approach is based on reweighting the important fragments, where a set of active reference structures are used to reweight the fragments in the reference structure. Secondly, a statistically supervised features selection and minifingerprint to select only the important fragments are applied. In this approach, searching is carried out by using sub-fragments that represent the important ones. Thirdly, a similarity coefficient based on mutually dependent fuzzy correlation coefficient is used. The last approach combined the best two out of the three approaches which are reweighting factors and fragment selection based on statistically supervised features selection. The proposed approaches were tested on the MDL Data Drug Report standard data set. The overall results of this research showed that the proposed fragment reweighting approaches outperformed the conventional industry-standard Tanimoto-based similarity search approach.