Friday 7th June, 2013
10:30am to 11:15am
Bayesian statistics use an alternative formulation of conditional probabilities consistent with the axioms of probability. Although Bayesian methods are not novel (1812) they have found increasing use in both scientific fields and common discourse. Quantitative Structure Activity Relationships (QSAR) is an enduring challenge in molecular modeling which has fallen into ill repute. Many criticisms have been described with the occasional remedy; over-fitting, domain specificity, interpretability, lack of prospective utility and finally human interpretability.
We have begun work on a new framework for QSAR that may ameliorate these deficiencies. Our method uses the concept of a background database for regularization and prospective utility without requiring large amounts of system specific knowledge. Molecular descriptors are restricted to fundamental physically meaningful molecular properties, and we use bayes factors for model selection and construction.
We have applied this framework to 4 ligand binding systems with data from actual pharmaceutical development programs. We show our framework can offer improvement over a simple null model, and can function with the information flow typical of a development program.
Sign in to add slides, notes or videos to this session