You may or may not be aware that version 10 of ACD/HNMR and CNMR Predictor now offer the option to generate predictions using a neural network algorithm.
For those of you who have followed ACD/Labs for years, you may observe this as an intriguing development.
Since the release of our NMR prediction packages many years ago, ACD/Labs has been strongly committed to a modified HOSE (Hierarchical Organization of Spherical Environments) code approach for NMR prediction.
The reasons for this commitment were obvious. Over the years, the developers have undertaken many internal validation studies and the only conclusions made were that HOSE code was clearly the better choice and thus warranted the most future development (one recent example).
So it was during the development of version 10 of ACD/Structure Elucidator that the inclusion of neural network algorithms was once again considered within ACD/Labs NMR software. In the beginning, the consideration of employing a neural network algorithm was based on an effort to try and enhance the speed of NMR predictions within the structure elucidation software. The process involves generating predictions on suggested structures and comparing them directly to the experimental data. Because neural network algorithms are known to perform 100s to 1000s of times faster than their HOSE code counterpart, research was conducted on this.
What was not expected, was a discovery made by the Structure Elucidator Project Leader, Kirill Blinov during his evaluation of the neural network algorithm performance. It turns out that Kirill’s neural network algorithms appeared to be outperforming our current (at the time) HOSE code implementation in version 9.0 in several areas.
As a result, the information gathered from this development work allowed us to improve and optimize our HOSE code algorithm for release in version 10 NMR predictors.
To make a long story short, what was the final result?
As of right now, the HOSE Code algorithm available in version 10 of ACD/Labs NMR predictors, still outperforms the neural network algorithm in all criteria except speed, hence it remains the default algorithm in the current version of ACD/HNMR and CNMR Predictors. This being said, the neural network algorithm is available in version 10 of the software for users to try out and evaluate on their own. It can be accessed through the Options menu and selecting Spectrum Plot…
Of course the ideal scenario is one where a user does not have to choose which algorithm they want to use. We are currently working on an intelligent hybrid approach that hopefully will be available in version 11. One example of a combined approach has been attempted elsewhere, but we are looking more deeply than this for an alternative solution that will result in improved prediction. Much work is also being done on atomic increment-based algorithms as well.
Lastly, it is important for me to point out a couple of limitations of the neural network algorithm:
- Neural net predictions are a black box. Therefore a "calculation protocol" which shows which experimental data was used in the prediction (available with HOSE code) is not available.
- Prediction training, while possible, has not proven to outperform HOSE code training.
For more information on our internal validation results of these algorithms, see this presentation from ENC 2007