Feature combinations networks with statistical QSAR models.pdf
We have developed a methodology for the interpretation of binary statistical models in terms of feature contribution to the outcome of the model’s prediction. The methodology remains agnostic to the learning algorithm and supplements the prediction of active/inactive with the elucidation of the model’s reason for the given prediction. The interpretation has been successfully applied to models for the prediction of Ames mutagenicity. The interpretation of in silico predictions and its use in knowledge mining is explored in this presentation.
Presented by Sam Webb at UK-QSAR Autumn meeting, Cambridge, UK; 30th September 2014.