NMRPredict:

		Modgraph Home
		NMRPredict Overview
		Carbon 13 NMR Prediction Overview

Neural Network Prediction

The traditional HOSE Code approach to carbon 13 prediction works very well if your query structure is well represented by reference spectra in the underlying database. However, when the "structural space" in the query molecule is not well represented by the database, the HOSE code method is likely to give wholly unreliable predictions.

As well as the HOSE code prediction approach, NMRPredict also uses a Neural Network algorithm, which is much more general and error tolerant than the HOSE code approach and is much more accurate at predicting shifts not found in the underlying database.

A good example is the molecule below. Atom 7 can only be predicted to one sphere by the HOSE code approach. The predicted HOSE code value is 145.5 ppm but the values from the database have a huge range from 90.3 ppm to 227.9 ppm. The Neural Network predicts the atom at 116.5 ppm, within 1 ppm of the observed value.

The Network used in NMRPredict it highly trained. It should be realised that a Neural Network itself has no intelligence. The intelligence of a network is determined by its structure and most importantly by the selection of the input features. Basic principles of NMR spectroscopy should be entered into this input layer and this has been done with NMRPredict. Examples of the intelligence added to the Network are described in more detail elsewhere.

The Neural Network prediction method in NMRPredict was developed during the mid 1990s in the group of Professor Robien at the University of Vienna. It was developed by V Purtuc and gave a very broad application range to cover general organic chemistry.

• to be general enough and be able to handle problems from basic organic chemistry to complex natural products
• to be able to handle "unusual" organic chemistry like organometallics
• to be solvent specific during prediction
• to be able to use stereochemical information not restricted by ring size, in the same way as could be done within the HOSE code technology
• to make sure that only interpolation and not extrapolation occured when making predictions

The Neural Network implemented into NMRPredict has been thoroughly tested and proved to be both reliable and accurate. The 4,000,000 assigned chemical shift values of the available 345,000 reference spectra can be predicted with an average deviation between experimental versus calculated of below 2.00 ppm. This includes compounds which are not well handled by traditional prediction programs, such as ferroces and chromium complexes.