Neural Network Predictions

An example of the neural network prediction of NMR chemical shifts for a natural product is illustrated in Figure 10.2-7 together with the calculations from other methods. This molecule was chosen as it had been discovered [47]... 

Two models of practical interest using quantum chemical parameters were developed by Clark et al. [26, 27]. Both studies were based on 1085 molecules and 36 descriptors calculated with the AMI method following structure optimization and electron density calculation. An initial set of descriptors was selected with a multiple linear regression model and further optimized by trial-and-error variation. The second study calculated a standard error of 0.56 for 1085 compounds and it also estimated the reliability of neural network prediction by analysis of the standard deviation error for an ensemble of 11 networks trained on different randomly selected subsets of the initial training set [27]. 

Rost, B., Casadio, R., and Fariselli, P. (1996). Refining neural network predictions for helical transmembrane proteins by dynamic programming. Intell. Syst. Mol. Biol. 4, 192-200. 

Joyce SJ, Osguthorpe DJ, Padgett JA et al. (1995) Neural network prediction of glass-transition temperatures from monomer structure. J Chem Soc Faraday Trans 91 2491-2496... 

Bohr, H.G. Neural Network Prediction of Protein Structures, Springer-Verlag. Inc., New York, NY. 2001. 

Table 8.7 Results obtained from building a neural network predictive model for the c/s-butadiene content in styrene-butadiene copolymers...

Table 7.4 Confusion Matrix from an Ensemble of 100 Trained Neural Networks Predicting Gene Family Target Activity...

Pedersen, A. G. Nielsen, H. (1997). Neural network prediction of translation initiation sites in eukaryotes perspectives for EST and genome analysis. Ismb 5,226-33. 

H. Yang, Z. Ring, Y. Briker, N. McLean, W. Friesen and C. Fairbridge, Neural Network prediction of cetane number and density of fuel from its chemical composition determined by FC and GC-MS, Fuel 81, 65-74 (2002). 

Svozil, D., Sevcik, J.G. and Kvasnicka, V. (1997). Neural Network Prediction of the Solvatochro-mic Polarity/Polarizability Parameter JI2. J.Chem.Inf.Comput.Sci., 37,338-342. 

A feature of this group of methods was an attempt by the authors to structurally restrict the explored temperature dependency of vapor pressure to the theoretically derived dependency. This motivation is clearly justified by the nonlinear dependency of vapor pressure from temperature, which cannot be easily captured with the pure linear regression approach. The use of nonlinear methods, however, can solve this problem. For example, neural networks predicted saturated vapor pressure for 352 hydrocarbons with an RMSE = 0.12 compared to an RMSE = 0.25 using a linear method with the same descriptors [112]. 

Serra, J.R., Jurs, P.C. and Kaiser, K.L.E. (2001) Linear regression and computational neural network prediction of Tetrahymena acute toxicity for aromatic compounds from molecular structure. Chem. Res. Toxicol., 14, 1535—1545. 

The problem of the absolute syllable duration can be easily solved by having the neural network predict this z-score instead of an absolute duration, an approach followed by [456]. This then frees the neural network from phonetic factors completely and allows it to use only prosodic features as input. 

W. Y. Choy, B. C. Sanctuary, and G. Zhu,/. Chem. Inf. Comput. Sd., 37,1086 (1997). Using Neural Network Predicted Secondary Structure Information in Automatic Protein NMR Assignment. 

D. Svozil, J. Pospichal, and V. Kvasnicka,/. Chem. Inf. Comput. Sci., 35,924 (1995). Neural Network Prediction of Carbcm-13 NMR Chemical Shifts of Alkanes. 

No chapter on modern chemometric methods would be complete without a mention of artificial neural networks (ANN). In a simple form these attempt to imitate the operation of neurons in the brain. Such networks have a number of linked layers of artificial neurons, including an input and an output layer (see Figure 8.13). The measured variables are presented to the input layer and are processed, by one or more intermediate ( hidden ) layers, to produce one or more outputs. For example, in inverse calibration, the inputs could be the absorbances at a number of wavelengths and the output could be the concentration of an analyte. The network is trained by an interactive procedure using a training set. Considering the example above, for each member of the training set the neural network predicts the concentration of the analyte. The discrepancy between the observed and predicted values... 

Silverman, D. C., Artificial Neural Network Predictions of Degradation of Nonmetallic Lining Materials from Laboratory Tests, Corrosion, Vol. 50, No. 6, 1994, pp. 411-418. 

The number of neurons in the hidden layer must be related to the number of available input data NDAT. Normally, a fraction of the available data is used for training (e.g., 80%), while the rest are used for validation of the neural network predictions. The maximum number of neurons in the hidden layer is thus given by... 

The evaluation of each neural network prediction efficiency has been accomplished using the sum of the squared errors of the validation and training procedures and also by graphic analysis. 

Cutting parameters Experimental D Neural network predicted D Absolute percentage error for NN... 

CLASH, Crest level assessment of coastal structures by fuU scale monitoring, neural network prediction and hazard analysis on permissible wave overtopping. Fifth Framework Program of the EU, Contract no. EVK3-CT-2001-00058. www.clash-eu.org. 

H. Verhaeghe, Neural network prediction of wave overtopping at coastal structures, PhD. thesis. University Gent, Belgium, ISBN 90-8578-018-7 (2005). 

Recent information on berm breakwaters has been described by Andersen. Only part of his research was included in the CLASH database and consequently in the Neural Network prediction method. He performed about 600 tests on reshaping berm breakwaters and some 60 on nonreshaping berm breakwaters (fixing the steep slopes by a steel net). The true nonreshaping Icelandic type of berm breakwaters with large rock classes, has not been tested and, therefore, his results might lead to an overestimation. 

Equations (15.7) and (15.8) for a simple rubble mound slope include a berm of 373 50 wide and a wave wall at the same level as the armor crest Ac — Rc- A little lower wave wall will hardly give larger overtopping, but no wave wall at all would certainly increase overtopping. Part of the overtopping waves will then penetrate through the crest armor. No formulae are present to cope with such a situation, unless the use of the Neural Network prediction method. 

---