Mnova Logo

13C NMR Prediction

Prediction of 13C NMR chemical shifts is carried out in Mnova NMRPredict  using two different procedures which are then combined by means of the so called ´Best´ prediction. This algorithm is used, in the light of a number of heuristic rules, to decide the final unified predicted chemical shift for every individual carbon atom

First, a database oriented chemical shift prediction is carried out. This is done with an extended HOSE code method  (Hierarchically Ordered Spherical of Environment) which consists of a one dimensional coding of the chemical environment of each carbon atom. Starting from the atom of interest, all atoms bonded directly to this atom (first sphere), over two bonds (second sphere) – and so on – are coded using characters which define atom types, bond types, ring closures, and spheres. The number of described spheres depends on the length of the code.

The HOSE code approach works very well for query structures that are well represented in the reference collection. If atoms can be predicted to three spheres or more, the prediction can be considered to be very reliable.

1

However, if the query structure is not well represented in the database and the atom can only be predicted to one or two spheres the prediction cannot be relied upon at all. Also the HOSE code approach exactly reproduces the contents of the reference database, including every error within that reference database.

In order to cope with these issues, NMRPredict 13C Prediction also uses a Neural Network algorithm which is more error tolerant than the HOSE code approach, giving more accurate results when the query atom is not represented in the database. Overall, if a carbon can be predicted at high spheres (i.e. number of shells ≥4), then the Best prediction method takes the value from the HOSE code procedure. However, if the predicted carbon is not well represented in the HOSE database, the chemical shift from the Neural Network algorithm will be used.

2

1H NMR Prediction

Prediction of 1H NMR spectra follows a similar approach to the case of 13C spectra. First, a prediction algorithm that is based on tabulated chemical shifts for classes of structures, corrected with additive contributions from neighboring functional groups or substructures, is carried out. For a given molecule, the appropriate substructures are automatically assigned following a hierarchical list. These substructures provide the base value of a final predicted chemical shift. Furthermore, a complementary prediction approach based upon partial atomic charges and steric interactions is also performed. This algorithm, named CHARGE, is a composite program made up of a neural network based approach for the one-, two- and three-bond substituent effects plus a theoretical calculation of the long range effects of substituents. This method requires first the generation of 3D conformers from a 2D structure so the individual spectra of all conformers are predicted.  Finally, an average predicted spectrum is calculated (employing a Boltzmann weighted average of the shifts calculated for all low-energy conformers).

1H NMR Best prediction analyses the individual chemical shifts from the two complementary methods and gives a single, unified predicted chemical shift.