Unknowns, predictions

Principal component regression is accomplished in two steps, a calibration step and an unknown prediction step. In the calibration step, concentrations of the constituent(s) to be quantitated in each calibration standard sample are assembled into a matrix, y, and mean-centered. Spectra of standards are measured, assembled into a matrix X, mean-centered, and then an SVD is performed. Calibration spectra are projected onto the d principal components (basis vectors) retained and are used to determine a vector of regression coefficients that can be then used to estimate the concentration of the calibrated constituent(s). 

Pelkonen O, Tolonen A, Korjamo T et al (2009) From known knowns to known unknowns predicting in vivo drug metabolites. Bioanalysis 1 393-414... 

Compared with linear and nonlinear regression methods, the advantage of ANN is its ability to correlate a nonlinear function without assumption about the form of this function beforehand. And the trained ANN can be used for unknown prediction. Therefore, ANN has been widely used in data processing of SAR. But if we use ANN solely, sometimes the results of prediction may be not very reliable. Experimental results indicate that some of the test samples predicted by ANN as optimal samples are really not true optimal samples. This is a typical example of so-called overfitting that makes the prediction results of trained ANN not reliable enough. Since the data files in many practical problems usually have strong noise and non-uniform sample point distribution, the overfitting problem may lead to more serious mistake in these practical problems. 

One of the early triumphs of the Mendeleef Periodic Table was the prediction of the properties of elements which were then unknown. Fifteen years before the discovery of germanium in 1886, Mendeleef had predicted that the element which he called ekasilicon would be discovered, and he had also correctly predicted many of its properties. In Table 1.8 his predicted properties are compared with the corresponding properties actually found for germanium. 

Prediction implies the generation of unknown properties. On the basis of example data, a model is established which is able to relate an object to its property. This model can then be used for predicting values for new data vectors. 

Finally, a model has to be tested using an independent data set with compounds yet completely unknown to the model the test set. The complete process of building a prediction model is depicted in Figure 10.1-1 as a flow chart. 

Molecular similarity is also useful in predicting molecular properties. Programs that predict properties from a database usually hrst search for compounds in the database that are similar to the unknown compound. The property of the unknown is probably close in value to the property for the known... 

There are numerous articles and references on computational research studies. If none exist for the task at hand, the researcher may have to guess which method to use based on its assumptions. It is then prudent to perform a short study to verify the method s accuracy before applying it to an unknown. When an expert predicts an error or best method without the benefit of prior related research, he or she should have a fair amount of knowledge about available options A savvy researcher must know the merits and drawbacks of various methods and software packages in order to make an informed choice. The bibliography at the end of this chapter lists sources for reviewing accuracy data. Appendix A of this book provides short reviews of many software packages. 

In principal, synthesis route prediction can be done from scratch based on molecular calculations. However, this is a very difficult task since there are so many possible side reactions and no automated method for predicting all possible products for a given set of reactants. With a large amount of work by an experienced chemist, this can be done but the difficulty involved makes it seldom justified over more traditional noncomputational methods. Ideally, known reactions should be used before attempting to develop unknown reactions. Also, the ability to suggest reasonable protective groups will make the reaction scheme more feasible. 

Much of the experimental work in chemistry deals with predicting or inferring properties of objects from measurements that are only indirectly related to the properties. For example, spectroscopic methods do not provide a measure of molecular stmcture directly, but, rather, indirecdy as a result of the effect of the relative location of atoms on the electronic environment in the molecule. That is, stmctural information is inferred from frequency shifts, band intensities, and fine stmcture. Many other types of properties are also studied by this indirect observation, eg, reactivity, elasticity, and permeabiHty, for which a priori theoretical models are unknown, imperfect, or too compHcated for practical use. Also, it is often desirable to predict a property even though that property is actually measurable. Examples are predicting the performance of a mechanical part by means of nondestmctive testing (qv) methods and predicting the biological activity of a pharmaceutical before it is synthesized. 

Cahbration is an important focus in analytical chemistry. It is the process that relates instmment responses to chemical concentrations. It consists of two basic steps estimation of the cahbration model parameters, and then prediction for new samples of unknown concentration. Cahbration refers to the step of the analytical process in Figure 2 where measurements are related to concentrations of chemical species or other chemical information. 

The flavone, isoflavone, and flavonol-type dyes owe their importance to the presence of an o-hydroxy carbonyl stmcture within the molecule. Positions 4 and 5 can chelate with different metallic salts to give colored, insoluble complexes. In other words, these dyes require a mordant in order to fix them onto the fiber. Perkin was able to predict the stmcture of unknown flavones by comparing the color of their complexes with the color of known complexes (70). For example, ferric chloride gives a green color with 5-hydroxyflavones and a brown one with 3-hydroxyflavones (71). 

---