The objective of the proposed work is to continue developing novel descriptors and informatics/statistical methods that will serve as the building blocks of evolving predictive ADME-toxicology, ADMET, tools, which are accurate, reliable and applicable across a wide diversity of chemistry. The two classes of descriptors being developed are semi-structure based properties from the membrane-interaction [Ml]- QSAR paradigm, and the universal 4D-fingerprint descriptors coming from the 4D-QSAR paradigm. Each class of descriptors has been shown to both enhance the quality of resultant predictive ADMET models when used with 'traditional' descriptors, as well as to lead to significant predictive ADMET models when the 'traditional' descriptors fail in model construction. Moreover, the MI-QSAR descriptors can provide mechanistic information regarding those ADMET properties which involve the interaction of a compound with the membrane of a cell as is the case, for example, for blood-brain barrier penetration. The 4D-fingerprints explicitly contain the 3D conformational distribution information of a molecule. Once computed these descriptors can be stored and subsequently used in any predictive ADMET application. The 4D-fingerpints also provide a general set of molecular features that allow the meaningful estimation of ligand-protein binding, including serum proteins, based on the concept that similar ligands bind in a similar fashion to a common protein. Cluster and categorical statistical methods are to be developed as 'preprocessors' in the building of reliable manifold models from structurally diverse training sets, and the subsequent virtual screening of structurally diverse compounds. Logistic regression, including integration of PLS, will be developed as a tool to optimize categorical ADMET models for classification endpoint data sets, as well as cases where continuous QSAR models of limited reliability might better be replaced by robust categorical models constructed from partitioning the continuous endpoint measures of the training set. These descriptors, statistical and informatics methods, and the predictive tools derived from them, should permit the efficient, reliable and robust prediction of multiple ADMET measures of an organic molecule. These tools may be especially helpful in streamlining pre-clinical drug discovery programs by providing virtual screening information that will early-on in a program eliminate drug-candidates with marginal ADMET property profiles. ? ? ?
Showing the most recent 10 out of 16 publications
