Nonlinear QSAR models

Carlsson L, Helgee EA, Boyer S. Interpretation of nonlinear QSAR models applied to Ames mutagenicity data. J Chem Inf Model 2009 49 2551-2558. 

Neural networks were originally designed as a model for the activity of the human brain. However, a computational neural network can be thought of simply as a nonlinear regression model when applied to QSAR studies. The computational neural network is a... 

Lucic, B., Nadramija, D., Basic, 1. andTrinajstic, N. (2003) Toward generating simpler QSAR models nonlinear multivariate regression versus several neural network ensembles and some related... 

In addition to nonlinear lipophilicity relationships for the transport and distribution of drugs, nonlinear relationships on molar refractivity are frequently observed in QSAR studies of enzyme inhibition data (provided that MR values are scaled by a factor of 0.1, as usual) [60,63,64,66-68]. Two such examples are given in Eq. (63) (Escherichia coli DHFR) and Eq. (64) (Lactobacillus casei DHFR) [101]. The differences between both models could be explained after the 3D structure of the enzyme became known. Whereas all substituents of a benzyl ring contribute to biological activities in E. coli DHFR, only the 3- and 4-substituents show up in the QSAR model for L. casei DHFR but not the 5-substituents. This results from a narrower binding pocket in L. casei DHFR a (3-branched leucine hinders the accommodation of 5-substituents, whereas a more flexible methionine in the same position of E. coli DHFR opens a wider binding pocket [101] ... 

Table 4.15 Examples of QSAR models for estimating bioconcentration in fish nonlinear correlations.

While the QMOD method is complex, the complexity was driven by the requirements of physical reality. QSAR models of binding affinity must have an inter-dependence between model and molecular pose. There must also be an interdependence between structural variation and predicted bound pose. The models, if they are at all physically sensible, must also encapsulate nonlinear effects such as size exclusion (as in Figure 2.3). Last, ligands with very different underlying scaffolds, as in Figure 2.10, must be modeled in a manner that can recognize their capability to bind the same pocket with similar affinities. 

---