A Neural Network Based Virtual Screening of Cytochrome P450 3A4 Inhibitors La´szlo´ Molna´randGyo¨rgyM.Keser´u´* Computer Assisted Drug Discovery, Gedeon Richter Ltd., PO Box 27, H-1475 Budapest, Hungary Received 29 August 2001; revised 29 October 2001; accepted 13 November 2001 Abstract—AvirtualscreeningtesttoidentifypotentialCP4503A4inhibitorshasbeendeveloped.Molecularstructuresofinhibitors and non-inhibitors available in the Genetest database were represented using 2D Unity fingerprints and a feedforward neural net- workwastrainedtoclassifymoleculesregardingtheirinhibitoryactivity.Validationtestsrevealedthatourneuralnetrecognizesat least89%of3A4inhibitorsandsuggestusingthismethodologyinourvirtualscreeningprotocol. # 2002ElsevierScienceLtd.All rights reserved. Cytochrome P450 3A4 is one of the major polymorphic izoenzymes responsible for the metabolism of almost 50% of known drugs in humans. Inhibitors of this iso- enzyme might cause drug–drug interactions because of decreasing the clearance of other drugs metabolized by 3A4. Early identification of potential 3A4 inhibitors is therefore needed to minimize the risk of clinically rele- vant interactions. In vitro HTS techniques can be used toscreenout3A4inhibitorsfrompreviouslysynthesized compound libraries. 1 In silico virtual screening, how- ever, has a pivotal role in the evaluation of virtual libraries. Since the high resolution X-ray structure of 3A4 is not available, structure based approaches are limited to the application of homology models. Although a high throughput docking (HTD) protocol could identify potential substrates we could not dis- criminate substrates and inhibitors by this technique. 2 Application of 3D QSAR models in virtual screening represents another option 3,4 that requires high through- put alignment (HTA) methodologies such as the recently launched Flex-S. 5 In addition to fitting uncer- tainties introduced by HTA, 3D-QSAR approaches sufferfromaspeedlimit,thatisthegenerationofthe3D structures for all of the compounds studied. Developing a 3D QSAR model we assume that binding modes of compounds used for the training set are similar. Con- sidering the relatively large binding pocket of 3A4 6 itis clear that substrates should have a relatively large con- formational degree of freedom within the active site (simultaneous binding of two molecules is also possi- ble 7 ). Since known 3A4 inhibitors represent a structu- rally diverse set of compounds it is likely that they bind 0960-894X/02/$ - see front matter # 2002 Elsevier Science Ltd. All rights reserved. PII: S0960-894X(01)00771-5 Bioorganic & Medicinal Chemistry Letters 12 (2002) 419–421 Figure 1. Distribution of 3A4 inhibitory indices calculated for inhibi- tors and non-inhibitors in the 3A4 test set. *Corresponding author. Fax: +36-1-4326002; e-mail: gy.keseru@ richter.hu