Optimization of Functional Group Prediction from Infrared Spectra Using Neural Networks

Abstract

In a large-scale effort, numerous parameters influencing the neural network interpretation of gas phase infrared spectra have been investigated. Predictions of the presence or absence of 26 different substructural entities were optimized by systematically observing the impact on functional group prediction accuracy for the following parameters: training duration, learning rate, momentum, sigmoidal discrimination and bias, spectral data reduction with four different methods, number of hidden nodes, individual instead of multioutput networks, size of the training set, noise level, and 12 different spectral preprocessing functions. The most promising approaches included constant monitoring of training progress with a 500 spectra cross-validation set, increasing the number of spectral examples in the training set from 511 to 2588, employing variance scaling, and using specialized instead of multioutput networks. An overall recognition accuracy of 93.8% for the presence and 95.7% for the absence of functionalities was achieved, while perfect prediction was reached for several present functional groups.

Clustering of infrared spectra with Kohonen networks

“…The recent use of multilayered feed forward neural networks for the interpretation of spectral data shows good promise: the internal representations used by these networks are non-linear and are built up by a learning process based on examples. After the seminal works by Robb and Munk [2][3], many researchers explored the possibilities of these networks for the interpretation of infrared spectra [4][5][6][7][8]. However, the learning method used is a supervised learning process, and relies upon an existing structural classification.…”

Application of a Pruning Algorithm To Optimize Artificial Neural Networks for Pharmaceutical Fingerprinting

“…Analysis of the input parameters selected by pruning methods is useful for interpreting the predicted results, for setting the boundaries of the fingerprinting region, and for introducing new parameters containing more information for making classifications more efficiently. The current approach can be extended for detection of localized fingerprinting regions and interpretation of results is neural network analysis of mass spectra 19 and infrared spectra 20,21 or for QSAR (quantitative structure-activity relationship) studies using, for example, "spectrumlike" representations of chemical structures. 22 The general idea of pruning can be also used for interpretation of more complex ANNs, such as the neural device proposed by Bashkin et al 23 There are certain advantages gained by decreasing the width of the fingerprint region.…”

Automatic amide I frequency selection for rapid quantification of protein secondary structure from Fourier transform infrared spectra of proteins

“…This input data compression technique has also been applied successfully by Klawun and Wilkins' [70] neural network analysis who were looking at the optimization of functional group predictions from infrared spectra using neural networks. With the exception of our automatic amide I frequency selection procedure, boxcar averaging is generally applied with a number of absorbance values to be replaced by their average ranging from 1 to 10 (Boxcar_1 to Boxcar_10).…”