Predicting Activity Cliffs: Can We Use Machine Learning for Special Cases?

• Activity Number: 595
• Type: Invited
• Date/Time: Thursday, August 4, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Section on Physical and Engineering Sciences
• Author(s): Rajarshi Guha*+
• Companies: National Institutes of Health
• Keywords: Cheminformatics ; Chemical biology ; Machine learning ; Structure-activity relationship ; Activity cliff ; Prediction

Abstract

The representation of SAR data in the form of landscapes and the identification of activity cliffs in such landscapes is well known. A number of approaches have been described to identifying activity cliffs, including several network based methods such as the SALI approach (JCIM, 2008, 48, 646-658). While a network representation of an SAR landscape moves away from the intuitive idea of rolling hills and steep gorges, it allows us to apply a variety of quantitative analyses. In this talk I will first provide a description of the Structure Activity Landscape and the network model derived from SALI. I will then discuss the view of activity cliffs as outliers and then describe investigations where we attempt to predict these activity cliffs using traditional machine learning methods. However, rather than predict the activities themselves, we instead derive a larger dataset by evaluating pairwise SALI values and train the model on the expanded dataset. We consider linear regression and random forest models, using assay data selected from the ChEMBL database. Our results indicate that this approach is able to predict activity cliffs with accuracies ranging from 70% to 82%.