Novel Structure—Activity Insights from Neural Network Models
Tariq A. Andrea · ACS symposium series · 1995
In regression QSAR, ligand/protein binding is a linear or parabolic function of ligands' physico-chemical properties. Due to the absence of higher order and cross-product terms, dependence of binding on one property is invariant to others. By comparison, neural networks are capable of delineating highly non-linear features. Stepwise regression of benzene sulfonylureas' binding to acetolactate synthases enzyme indicates that, properties of the ortho substituent R 2 have the following order of significance: MR > π > F . Affinity depends parabolically on MR (MR opt =14.05) and increases linearly with π and F . The constant curvature (-0.0158) of the MR parabola indicates that the binding pocket tolerates R 2 's with 4-7 heavy atoms and that this tolerance is invariant to π and F . This suggests an enzyme pocket with fixed size. Neural networks analysis finds the same order of significance of R 2 properties. While in this model MR dependence is not mathematically parabolic, it has a "parabolic or Guassian-like" shape. Like regression QSAR, the neural model indicates that optimal MR and tolerance to size variation depends on π and F . It suggests a binding pocket which accommodates larger hydrophobic and electron withdrawing substituents than hydrophilic and electron donating ones. It also indicates higher tolerance to size variations in hydrophobic and electron withdrawing substituents than in hydrophilic and electron donating ones. These are consistent with x-ray crystallographic findings that even structurally related ligands can bind differently to the same protein.