Ligand-based models are often used in drug discovery to select compounds for further study. For example, they may be used in virtual screens to identify novel inhibitors, or they may be used in lead optimization to discard compounds with poor physical properties. Such models are built and evaluated using statistical techniques based on available data. Typically this data comes from literature sources or prior screening campaigns and the models are applied to novel chemistries. Unfortunately, this process breaks the key requirement of most statistical modeling techniques that the training and testing distributions are identical. In this talk we explore the extent to which this requirement holds in typical ligand-based modeling efforts. We examine the implications of its failing to hold and share a well tested general modeling approach that overcomes these problems.

Reducing the bias in ligand-based modelling

244th National Meeting (2012)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.