Building and Model Selection

Iteration of the steps, descriptor selection, model building, and model validation in combination with an optimi ation algorithm allows one to select a descriptor subset having maximum predictivity. 

Nevertheless, chemists have been planning their reactions for more than a century now, and each day they run hundreds of thousands of reactions with high degrees of selectivity and yield. The secret to success lies in the fact that chemists can build on a vast body of experience accumulated over more than a hundred years of performing millions of chemical reactions under carefully controlled conditions. Series of experiments were analyzed for the essential features determining the course of a reaction, and models were built to order the observations into a conceptual framework that could be used to make predictions by analogy. Furthermore, careful experiments were planned to analyze the individual steps of a reaction so as to elucidate its mechanism. 

Obtaining Kinetic Samples for Reactive Extrusion. To develop and test kinetic models, homogeneous samples with a well defined temperature-time history are required. Temperature history does not necessarily need to be isothermal. In fact, well defined nonisothermal histories can provide very good test data for models. However, isothermal data is very desirable at the initial stages of model building to simplify both model selection and parameter estimation problems. 

This section will provide an overview on ADME models from our group to illustrate our approach for building predictive models on structurally diverse training sets. Datasets for intestinal human absorption and human serum albumin binding are discussed, while models for other relevant ADME properties have also been obtained. Those models, however, do not stand alone but are used in combination with those models tailored for affinity and selectivity in the frame of multidimensional lead optimization. 

Because onh a few variables are selected to build the models, MLR begins to approach the univariate methods. Tliis is especially limiting during prediction where there is little validation of the results. MLR is also limited to relatively simple systems (i.e., small number of components and other sources of variation) and does not lu e ihe full multivariate advantage. Tlie main advantage of MLR is its simplicity—the final models are easy to explain to other team members. 

To illustrate the MLR method, the SMLR calibration method is used to build a model for the czs-butadiene content in the polymers. In this case, four variables are specified for selection, based on prior knowledge that there are four major chemical components that are varying independently in the calibration samples. The SMLR method chooses the four X-variables 1706, 1824, 1670, and 1570 nm, in that order. These four selected variables are then used to build an MLR regression model for czs-butadiene content, the fit of which is shown in Figure 8.13. Table 8.5 lists the variables that were chosen by the SMLR method,... 

---