Spectrum prediction neural networks

Applications of neural networks are becoming more diverse in chemistry [31-40]. Some typical applications include predicting chemical reactivity, acid strength in oxides, protein structure determination, quantitative structure property relationship (QSPR), fluid property relationships, classification of molecular spectra, group contribution, spectroscopy analysis, etc. The results reported in these areas are very encouraging and are demonstrative of the wide spectrum of applications and interest in this area. 

We have already met one tool that can be used to investigate the links that exist among data items. When the features of a pattern, such as the infrared absorption spectrum of a sample, and information about the class to which it belongs, such as the presence in the molecule of a particular functional group, are known, feedforward neural networks can create a computational model that allows the class to be predicted from the spectrum. These networks might be effective tools to predict suitable protective glove material from a knowledge of molecular structure, but they cannot be used if the classes to which samples in the database are unknown because, in that case, a conventional neural network cannot be trained. 

Ergungor describes the application of on-line Raman spectroscopy and neural networks to the simultaneous prediction of temperature and crystallinity of nylon-6 nanocomposites as a function of cooling rate. The authors prefer their neural network approach because they make use of information in the entire spectrum rather than from a few bands as most studies have done.84 Van Wijk etal. of Akzo Nobel obtained a patent on the use of a Raman spectrum of a polymeric fiber to determine dye uptake and other structural or mechanical properties based on previously developed models.85... 

The class type dehnes the keyword for a class vector for neural network training and prediction. Depending on the number of components of the class either a multicomponent class vector (e.g., a spectrum) or a single-component class property is allocated. The number of components automatically dehnes the number of weights used in the ANNs. 

Artificial neural networks do not require any information about the relationship between spectral features and corresponding substructures in advance. The lack of information about complex effects in a vibrational spectrum (e.g., skeletal and harmonic vibrations, combination bands) does not affect the quality of a prediction or simulation performed by a neural network. 

FIGURE 6.5 The infrared spectrum of a query compound compressed by Hadamard transform for the prediction of benzene derivatives by a CPG neural networks. The spectrum exhibits some typical bands for aromatic systems and chlorine atoms. 

FIGURE 6.8 Prediction of a bicyclic compound by CPG neural networks (low-pass D20 Cartesian RDF, 128 components). Eight best matching structures from the descriptor database and the RMS errors between their descriptor and the one predicted from the Kohonen network. The structure belonging to the query spectrum was found at the lowest RMS error of 0.122. 

Database Approach is a specific method for deriving the molecular structure from an infrared spectrum by predicting a molecular descriptor from an artificial neural network and retrieving the structure with the most similar descriptor from a structure database. 

Spectrum Simulation is a method for creating spectra from information about the chemical structure of a molecule, for which none exist this is typically supported by prediction technologies, such as artificial neural networks. 

Spectrum prediction is a frequently used technique during the structure elucidation process, but a detailed inspection of the results is necessary. Some programs offer the possibility to use different algorithms for spectrum prediction (usually HOSE code technology and neural networks). In such a situation both methods should be applied and the results obtained should be carefully compared [22]. At least in the case of different predictions a further critical evaluation of the result should be an obligation. 

Spectrum Prediction with Artificial Neural Networks... 

As mentioned above neural networks learn about the correlation between the structure and the IR spectrum by examining a set of molecules and their corresponding IR spectra. Once a network has been trained with a set of examples it is able to predict an IR spectrum for a compound it has not seen before. 

A totally different approach has been reported by Steinhauer et al. The IR spectrum is used as input into a CPG neural network. The network supplies the corresponding radial code (see Section 4.3) as output. This simulated radial code can be iteratively decoded to the molecular 3D structure. This method allows one, for the first time, to obtain 3D information from an IR spectrum. An example, the prediction of the 3D structure of diphenyl ether, is shown in Figure 9. 

---