QSAR approaches that build regression equations expressing the relationship between molecular structural features and ligand activities are purely correlative methods. The Surflex-QMOD approach adopts a different strategy, attempting to construct causal models that can quantitatively predict both ligand activity and binding mode to targets where sets of ligands bind competitively and where activity is related to binding free-energy. The approach uses multiple-instance machine-learning to evolve a physical binding site model at the same time as the ligand binding modes are elucidated. We present results in two areas: 1) where the process of modeling proceeds iteratively, with QMOD models being successively refined in order to make predictions on future compounds; and 2) where the model-building benefits quantitatively from biophysical information in the form of protein-ligand complexes. QMOD binding site models will be shown to be more predictive than other approaches when scaffold diversity is high, and especially when predictions are made based on past compounds in order to predict future ones. Correlative models often appear to work in traditional QSAR assessments like cross-validation, but models that move toward the causal basis for molecular activity make better predictions where new molecules diverge from those that are known.

Advances in causal QSAR modeling: Structural guidance

248th National Meeting (2014)

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