The efficacy of neural network (NN) and partial least-squares (PLS) methods is compared for the prediction of NMR chemical shifts for both 1H and 13C nuclei using very large databases containing millions of chemical shifts. The chemical structure description scheme used in this work is based on individual atoms rather than functional groups. The performances of each of the methods were optimized in a systematic manner described in this work. Both of the methods, least-squares and neural network analyses, produce results of a very similar quality, but the least-squares algorithm is approximately 2--3 times faster.

BackgroundThis article coincides with the 40 year anniversary of the first published works devoted to the creation of algorithms for computer-aided structure elucidation (CASE). The general principles on which CASE methods are based will be reviewed and the present state of the art in this field will be described using, as an example, the expert system Structure Elucidator.ResultsThe developers of CASE systems have been forced to overcome many obstacles hindering the development of a software application capable of drastically reducing the time and effort required to determine the structures of newly isolated organic compounds. Large complex molecules of up to 100 or more skeletal atoms with topological peculiarity can be quickly identified using the expert system Structure Elucidator based on spectral data. Logical analysis of 2D NMR data frequently allows for the detection of the presence of COSY and HMBC correlations of "nonstandard" length. Fuzzy structure generation provides a possibility to obtain the correct solution even in those cases when an unknown number of nonstandard correlations of unknown length are present in the spectra. The relative stereochemistry of big rigid molecules containing many stereocenters can be determined using the StrucEluc system and NOESY/ROESY 2D NMR data for this purpose.ConclusionThe StrucEluc system continues to be developed in order to expand the general applicability, provide improved workflows, usability of the system and increased reliability of the results. It is expected that expert systems similar to that described in this paper will receive increasing acceptance in the next decade and will ultimately be integrated directly to analytical instruments for the purpose of organic analysis. Work in this direction is in progress. In spite of the fact that many difficulties have already been overcome to deliver on the spectroscopist's dream of "fully automated structure elucidation" there is still work to do. Nevertheless, as the efficiency of expert systems is enhanced the solution of increasingly complex structural problems will be achievable.

The accuracy of (13)C chemical shift prediction by both DFT GIAO quantum-mechanical (QM) and empirical methods was compared using 205 structures for which experimental and QM-calculated chemical shifts were published in the literature. For these structures, (13)C chemical shifts were calculated using HOSE code and neural network (NN) algorithms developed within our laboratory. In total, 2531 chemical shifts were analyzed and statistically processed. It has been shown that, in general, QM methods are capable of providing similar but inferior accuracy to the empirical approaches, but quite frequently they give larger mean average error values. For the structural set examined in this work, the following mean absolute errors (MAEs) were found: MAE(HOSE) = 1.58 ppm, MAE(NN) = 1.91 ppm and MAE(QM) = 3.29 ppm. A strategy of combined application of both the empirical and DFT GIAO approaches is suggested. The strategy could provide a synergistic effect if the advantages intrinsic to each method are exploited.

No abstract

The validation of the performance of a neural network based 13C NMR prediction algorithm using a test set available from an open source publicly available database, NMRShiftDB, is described. The validation was performed using a version of the database containing ca. 214,000 chemical shifts as well as for two subsets of the database to compare performance when overlap with the training set is taken into account. The first subset contained ca. 93,000 chemical shifts that were absent from the ACD\CNMR DB, the "excluded shift set" used for training of the neural network and the ACD\CNMR prediction algorithm, while the second contained ca. 121,000 shifts that were present in the ACD\CNMR DB training set, the "included shift set". This work has shown that the mean error between experimental and predicted shifts for the entire database is 1.59 ppm, while the mean deviation for the subset with included shifts is 1.47 and 1.74 ppm for excluded shifts. Since similar work has been reported online for another algorithm we compared the results with the errors determined using Robien's CNMR Neural Network Predictor using the entire NMRShiftDB for program validation.