Free energy perturbation and Free-Wilson models compared and contrasted.
Martin Packer1
1AstraZeneca
An additive group contribution model for binding free energy was proposed by Spencer Free and James Wilson in 1964: and this Free-Wilson model for free energy continues to be a valuable tool in drug design. The Free-Wilson model was an early attempt to formulate mathematical approaches to prediction of affinity and structure activity relationships (SAR), with the additional benefit that R-groups could be enumerated to suggest molecules as yet untested or unsynthesised. Ten years before publication of the Free-Wilson model, Zwanzig outlined a statistical mechanical approach to prediction of free energy changes, called free energy perturbation (FEP). This approach is applicable to any system for which we can sample states using a high quality potential function. Adoption of FEP methodology to predict protein-ligand binding affinity took several decades, not least because the first usuable protein-ligand structures emerged only in the 1970’s. Enhanced sampling algorithms and computer hardware have made FEP generally tractable only within the last decade, at least for use in drug design, but there are now multiple published case studies of its application to drug design.
In this talk I will explore the extent to which simple Free-Wilson models and complex FEP models can complement one another. I will discuss data sets in which the models show very similar trends and examples where the models diverge significantly. When free energy additivity fails to account for SAR trends in a Free-Wilson model, it is often assumed that R-group co-operativity is the root cause, so that groups bind with different affinity in different ligands. FEP should be able to avoid this issue, but it is important that we assess whether non-additive SAR trends are valid. FEP, as a physics-based approach, has been shown to deliver high accuracy in prediction of affinity and such models should play an increasing role in drug design, especially if the trend towards automated library design continues. This will demand that we deploy a range of robust approaches for assessment of model error and consistency of SAR trends, using fundamental models such as the Free-Wilson approach.
