ENC 2008

March 9–14, 2008

Asilomar Conference Grounds
Pacific Grove, CA, USA

Conference Details

Hospitality Suite: Oak Shelter

ACD/Labs Seminar

ACD/Labs NMR Software Symposium at ENC 2008. March 9, 2008.

Poster Schedule

Monday, March 10 & Tuesday, March 11, 2008
Time: 2:00–4:00 PM (Author will be available at poster)

TITLE: An Efficient Incremental Scheme for ¹⁵N, ¹⁹F, and ³¹P Chemical Shift Prediction
Kirill A. Blinov, Yegor D. Smurnyy, Tatiana S. Churanova, Mikhail E. Elyashberg (ACD Ltd.), Brent A. Lefebvre (ACD/Labs), Antony J. Williams (ChemZoo)

Abstract: The implementation of an atomic increments algorithm for the prediction of ¹H and ¹³C chemical shifts has been previously reported [1]. Here, we extend this algorithm to ¹⁵N, ¹⁹F, and ³¹P nuclei, evaluating the quality of different incremental schemes. Non-linear regression was performed either by partial least squares (PLS) or neural network (NN) algorithms. A standard database-based HOSE-code approach was also applied to the same nuclei for performance evaluation purposes.

Unlike ¹³C and ¹H, the number of experimentally measured chemical shifts for ³¹P, ¹⁹F, and especially ¹⁵N is low. Currently, we have collected ~66,000 chemical shift values for ¹⁹F, ~40,000 for ³¹P, and ~25,000 for ¹⁵N. Therefore the methods of structure encoding and neural network training previously used for ¹H and ¹³C chemical shift prediction need to be further adapted to these heteronuclei. Previously [2], we presented the atomic increments method based on both individual atomic descriptors and their combinations, the so-called cross-increments, that help evaluate synergistic effects between substituents. In a typical ¹H or ¹³C calculation, the total number of independent variables (the regular and cross-increments combined) was between 5000 and 10,000. For heteronuclei, we have found that such a detailed scheme leads to overtraining and poor results for structures that are not included in the training set. We systematically tested different schemes, varying the number of individual atomic classes and the extent of the atomic neighborhood taken into account. Standard errors of 9 ppm (¹⁵N), 5 ppm (¹⁹F), and 7 ppm (³¹P) (the best results) were obtained for a configuration with 15–25 atomic classes that considered only atoms separated from the heteronucleus by no more than 5 covalent bonds, resulting in 1000 to 2000 individual independent variables. This compares to 70 atomic classes for ¹³C, illustrating the need for a more general description scheme in the case of limited datasets. In all cases, the scheme outperformed the HOSE code approach.

Surprisingly, the quality of the results did not depend on the non-linear regression algorithm. The highly popular neural networks, and somewhat overseen and older PLS method, both performed nearly identically. We conclude that the modern computation power of today's computers allows the best-fit solution to be found with either regression method. This means that the residual error must arise from three other sources: a) an inadequate description of the chemical structure; b) variability of experimental conditions that affect the chemical shift, such as solvent and temperature; and c) errors in the training set, mostly due to the human factor. Work currently underway will address the first two issues by improving the description scheme that will take into account extended conjugation, sterical effects, and conditions at which the data were acquired.

[1] Toward More Reliable ¹³C and ¹H Chemical Shift Prediction: A Systematic Comparison of Neural-Network- and Least-Squares-Regression-Based Approaches. Y.D. Smurnyy, K.A. Blinov, T.S. Churanova, M.E. Elyashberg, A.J. Williams J.; Chem. Inf. Model. 2007 (in press).
[2] NMR Chemical Shift Prediction by Atomic Increment-Based Algorithms. Y.D. Smurnyy, K.A. Blinov, and M.E. Elyashberg, B.A. Lefebvre, and A.J. Williams; ENC 2007 Poster Session.

Wednesday, March 12 & Thursday, March 13, 2008
Time: 2:00–4:00 PM (Author will be available at poster)

TITLE: Validating the Quality of Large Collections of NMR Spectra Automatically
Sergey S. Golotvin, Eugene Vodopianov, Rostislav Pol, Mikhail Kvasha (ACD Ltd.), Brent A. Lefebvre (ACD/Labs), Jim Brien, Chris Lein, and Tom Moertl (Sigma-Aldrich Corp.)

Abstract: As NMR instrumentation continues to advance, more and more spectral information becomes available for researchers with a finite amount of time. A typical ¹H–¹³C HSQC experiment for example, can now be performed in ten minutes or less, making it feasible to employ this spectrum for high-throughput evaluation of a proposed chemical structure. Although structure evaluation methods utilizing the combination of ¹H and ¹H–¹³C HSQC data were shown to offer significantly higher accuracy [1,2], an evaluation based only on a 1D ¹H spectrum is still desirable in many instances.

In the work presented here, 15,000 ¹H NMR spectra from the Aldrich collection were evaluated in complete automation using a number of proton prediction methods [3]. The most accurate prediction method was able to confirm 88% of all structures as consistent, and flagged less than 5% as inconsistent. A manual examination of these flagged spectra did indeed reveal some amount of truly wrong structures as well as shortcomings in prediction, processing, and analysis steps. A discussion of factors limiting the verification performance is given.

[1] Towards the automatic analysis of NMR spectra: Part 6. Confirmation of chemical structure employing both ¹H and ¹³C NMR spectra, Griffiths L, Horton R. Magn. Reson. Chem. 2005; 44: 139.
[2] Automated structure verification based on combination of 1D ¹H NMR and 2D ¹H–¹³C HSQC spectra, Golotvin SS, Vodopianov E, Rostislav P, Lefebvre BA, Williams AJ, Rutkowske RD and Spitzer TD, Magn. Reson. Chem. 2007; 45, 803–813.
[3] Automated structure verification based on ¹H NMR prediction, Golotvin SS, Vodopianov E, Lefebvre BA, Williams AJ, Spitzer TD. Magn. Reson. Chem. 2006; 44, 524–538.

For complete event information, visit ENC.