Selecting Screening Candidates for Gene Family Targets Using Neural Networks and Statistical Methods

David J. Livingstone^1,2 *, Will R. Pitt³, Emanuela Gancia³, John G. Montana⁴, Martyn G. Ford², David C. Whitley², David T. Manallack⁵

¹ChemQuest, Delamere House, 1 Royal Crescent, Sandown, Isle of Wight, PO36 8LZ UK

²Centre for Molecular Design, University of Portsmouth, King Henry Building, King Henry I Street, Portsmouth, Hampshire, PO1 2DY, UK

³Celltech R&D Ltd., Granta Park, Great Abington, Cambridge, CB1 6GS, UK

⁴Amedis Pharmaceuticals Unit 209, Cambridge Science Park, Milton Road, Cambridge CB4 0GZ, UK

⁵De Novo Pharmaceuticals, Compass House, Vision Park, Histon, Cambridge, CB4 9ZR

Modern techniques in drug research, such as combinatorial chemistry and high throughput screening, are aimed at increasing the rate at which new “hits” are discovered and, therefore, the speed of delivery of new drugs to market. One of the problems in constructing combinatorial libraries or selecting compounds for purchase is the question of “drug likeness”. Much research has been directed at the creation of In Silico methods for the prediction of ADMET properties in order to avoid progressing otherwise promising candidates which will ultimately fail in development.

In addition to drug likeness, of course, there is also the matter of activity at a desired biological target. The MDDR database was used to provide compounds targeted against gene families and sets of randomly selected molecules. BCUT parameters were computed in order to characterize the structures. A series of neural networks, trained using consensus methods, identified over 80% of the compounds targeting a gene family. A number of other mathematical modeling methods including Bayesian neural networks have been used to study this problem and the results will be compared and discussed.