Structure/property estimation method

Classes of Estimation Methods Table 1.1.1 summarizes the property estimation methods considered in this book. Quantitative property-property relationships (QPPRs) are defined as mathematical relationships that relate the query property to one or several properties. QPPRs are derived theoretically using physicochemical principles or empirically using experimental data and statistical techniques. By contrast, quantitative structure-property relationships (QSPRs) relate the molecular structure to numerical values indicating physicochemical properties. Since the molecular structure is an inherently qualitative attribute, structural information has first to be expressed as a numerical values, termed molecular descriptors or indicators before correlations can be evaluated. Molecular descriptors are derived from the compound structure (i.e., the molecular graph), using structural information, fundamental or empirical physicochemical constants and relationships, and stereochemcial principles. The molecular mass is an example of a molecular descriptor. It is derived from the molecular structure and the atomic masses of the atoms contained in the molecule. An important chemical principle involved in property estimation is structural similarity. The fundamental notion is that the property of a compound depends on its structure and that similar chemical stuctures (similarity appropriately defined) behave similarly in similar environments. 

In another example, Chavali et al. demonstrated that 2D connectivity indices can give good structure/property correlations in molybdenum-catalyzed epoxidation [53,54]. They used the Computer Aided Molecular Design (CAMD) environment, a powerful computational tool used in product design. The method uses optimization techniques coupled with molecular design and property estimation methods, generating those molecular structures that match a desired set of properties. 

Physical property estimation methods may be classified into six general areas (1) theory and empirical extension of theory, (2) corresponding states, (3) group contributions, (4) computational chemistry, (5) empirical and quantitative structure property relations (QSPR) correlations, and (6) molecular simulation. A quick overview of each class is given below to provide context for the methods and to define the general assumptions, accuracies, and limitations inherent in each. 

Molecular structure design relies on accurate property estimation methods. When sufficiently accurate, the atoms and groups in the molecular structure are adjusted to minimize the sum of the squares of the differences between the property estimates and the specified values ... 

An extensive series of studies for the prediction of aqueous solubility has been reported in the literature, as summarized by Lipinski et al. [15] and jorgensen and Duffy [16]. These methods can be categorized into three types 1 correlation of solubility with experimentally determined physicochemical properties such as melting point and molecular volume 2) estimation of solubility by group contribution methods and 3) correlation of solubility with descriptors derived from the molecular structure by computational methods. The third approach has been proven to be particularly successful for the prediction of solubility because it does not need experimental descriptors and can therefore be applied to collections of virtual compounds also. 

Group contribution techniques are based on the concept that a particular physical property of a compound can be considered to be made up of contributions from the constituent atoms, groups, and bonds the contributions being determined from experimental data. They provide the designer with simple, convenient, methods for physical property estimation requiring only a knowledge of the structural formula of the compound. 

Methods have been presented, with examples, for obtaining quantitative structure-property relationships for alternating conjugated and cross-conjugated dienes and polyenes, and for adjacent dienes and polyenes. The examples include chemical reactivities, chemical properties and physical properties. A method of estimating electrical effect substituent constants for dienyl and polyenyl substituents has been described. The nature of these substituents has been discussed, but unfortunately the discussion is very largely based on estimated values. A full understanding of structural effects on dienyl and polyenyl systems awaits much further experimental study. It would be particularly useful to have more chemical reactivity studies on their substituent effects, and it would be especially helpful if chemical reactivity studies on the transmission of electrical effects in adjacent multiply doubly bonded systems were available. Only further experimental work will show how valid our estimates and predictions are. 

Fig. 10. Estimated viscoelatic properties in a normal human breast in vivo. (A) T2 anatomical image. (B) Shear modulus image of the same slice. (Q Young s modulus image of the same slice. Grey scale bars are in kPa. Images B and C are extracted from 3D data sets of reconstructed elasticity parameters, obtained with the subzone based method used in Fig. 8. Note the good contrast in image C, even though the mechanical parameters are not obviously correlated to the structural properties depicted in image A (reprinted with permission from Ref. 48 2000 IOP Publishing Ltd.).

There are many methods to estimate physico-chemical properties (a. 18 and a. 19), and calculate these using thermodynamic and empirical relationships (a.20). It is particularly important where estimation methods have been used to ensure that the results derived are consistent with one another and are reasonable based on chemical structure. 

Quantitative Stmcture-Activity Relationships (QSARs) are estimation methods developed and used in order to predict certain effects or properties of chemical substances, which are primarily based on the structure of the substance. They have been developed on the basis of experimental data on model substances. Quantitative predictions are usually in the form of a regression equation and would thus predict dose-response data as part of a QSAR assessment. QSAR models are available in the open literature for a wide range of endpoints, which are required for a hazard assessment, including several toxicological endpoints. 

When theoretical understanding is insufficient and quantitative correlations are not available, we can often make useful qualitative estimations by using fragmentary empirical structure-property relations. The principal tools are observations of associations and trends, which are often the only methods available in biological, health, safety, and environmental properties. 

To fill the data gaps for determining potential hazards to human health and the environment, the EPA often relies on internally developed estimation methods. These include empirical data available for structural analogs and computational methods for the estimation of physicochemical properties, which in turn are used to estimate environmental fate, bioavailability, toxicity in humans and aquatic organisms, and exposure [17]. 

Therefore, similar to the attempts made to estimate vapor pressure (Section 4.4) there have been a series of quite promising approaches to derive topological, geometric, and electronic molecular descriptors for prediction of aqueous activity coefficients from chemical structure (e.g., Mitchell and Jurs, 1998 Huibers and Katritzky, 1998). The advantage of such quantitative structure property relationships (QSPRs) is, of course, that they can be applied to any compound for which the structure is known. The disadvantages are that these methods require sophisticated computer software, and that they are not very transparent for the user. Furthermore, at the present stage, it remains to be seen how good the actual predictive capabilities of these QSPRs are. 

Clearly, this example demonstrates how important it is to recognize the structural difference between similar compounds and base property estimation on AStructure-ATm relationships instead of simply setting their Tm values equal to each other. Figures 10.4.2 to 10.4.6 illustrate similarity-based estimation of Tm using the method of Joback and Reid (Section 9.3). For comparison, the observed Tm values [4] for the query compounds are given below ... 

The PMN review process has evolved over time within the constraints set by TSCA. An important constraint is that submitters are required to furnish only test data already in their possession (if any) and are not required to conduct a battery of tests as a precondition for approval. This generalization holds true for basic chemical property data as well as toxicity data, and it is the main reason why TSCA has been such a powerful impetus for developing estimation methods for many of the parameters needed in environmental assessment. To illustrate how extreme the situation is, in one study of more than 8,000 PMNs for class 1 chemical substances (i.e., those for which a specific chemical structure can be drawn) that were received from 1979 through 1990,Lynch et al. (1991) found only 300 that contained any of the property data noted earlier as needed for environmental assessment. The U.S. is unique among industrialized nations in requiring its assessors to work in the virtual absence of test data. 

As in Lyman s Handbook, emphasis is on broadly applicable estimation methods. Given the many and varied reasons that one might be interested in chemical property estimation, we believe that most users of this book will have less interest in chemical class-specific estimation methods. Obviously such methods are reliable only for that class, which may be defined very narrowly, and they may produce substantial yet unknown error if applied to compounds that differ significantly. Many of the newer methods were developed using much larger and more varied training sets, thus are more likely to be useful for diverse and/or structurally complex compounds. Therefore, in contrast to the situation that existed in 1982 when Lyman s Handbook was published, current users often do not need to make decisions about which of several class-specific methods seems most applicable to the compound of interest. 

As originally conceived, this handbook was to include worked examples of estimation methods for a group of benchmark chemicals for which reliable properties exist. The advantage of this approach is that the reader is likely to find it easier to apply the estimation methods if there are examples to follow. This proved to be more difficult than was expected and not all these benchmark chemicals are fully treated. Ideally, the estimated values should correspond closely with measured values. In some cases there are significant discrepancies, and this serves to reinforce the message that there remains a need to improve these methods in both accuracy and scope. The subject of estimating chemical properties from molecular structure and from related properties is thus a fruitful topic of research, and will remain so for many years into the future. 

This relationship identifies enthalpy and entropy as the thermodynamic properties that determine the value of the melting point. Because each of these properties shows a different dependence upon structure, attempts to estimate the melting point of a wide variety of organic compounds have been largely unsuccessful. The next two sections seek to correlate AHm and ASm with aspects of molecular structure and present methods for their estimation. 

With one exception, the estimation methods considered for this book are those that use only structural information as input, since that is all that is likely to be available i.e., if no boiling point is available for a chemical, then it is unlikely that other property values will be available from which a boiling point could be estimated (via some property-property correlation). The one exception is for estimation methods using melting points as inputs, since a value for a melting point (but not a boiling point), especially for thermally unstable chemicals, is likely to be available in the literature. 

A test set of 6 to 13 aroma compound partition coefficients between different food contact polymers (low density polyethylene (LDPE), high density polyethylene (HDPE) polypropylene (PP), polyethylene terephthalate (PET), polyamide (PA)) and different food simulant phases (water, ethanol, aqueous ethanol/water mixtures, methanol, 1-propanol) were taken from the literature (Koszinowski and Piringer, 1989, Baner, 1992, Franz, 1990, Koszinowski, 1986, Franz, 1991, Baner, 1993, Piringer, 1992). Table 4-2 shows the test set of 13 different aroma compounds, with their properties and their structures. The experimental data were compared to estimations using different estimation methods of UNIFAC-FV, GCFLORY (1990), GCFLORY (1994) and ELBRO-FV. 

Estimation of log P by using quantitative structure property relationships (QSPR) modeling and molecular descriptors (described above) has resulted in a number of highly accurate methods. Methods involving MLR, PLS, and artificial neural network ensembles (ANNE) modeling have been reviewed.In summary, estimation of partition coefficient has now reached a stage where the error associated with estimation is approximately equal to experimental error and reliable estimates can be obtained in silico. 

---