Items of interest for the JEOL community

Krish Krishnamurthy

Above: Krish Krishnamurthy pictured, most likely discussing his favorite topic.

NMR data contains a trove of useful information for answering a wide variety of chemical and biological problems. However, with this broad utility comes complexity. Converting an NMR spectrum into useable information is a challenge because the workflow for NMR data analysis is primarily based on manual processing and interpretation of each individual spectrum. This can be a labour-intensive process which quickly becomes unfeasible as the number of samples to analyse increases, and if the data are to be compared across samples, then the information must be distilled into tabular or numerical format to facilitate such comparisons. These needs for throughput, efficiency, and data simplification led to research in automated computational methods for NMR analysis. However, automated attempts at extracting information from Fourier transformed NMR spectra, such as chemical shifts of peaks, integrals, peak heights, etc., quickly ran into difficulties. Some of the major issues arose from artifacts introduced during the Fourier transform, including baseline issues and phase distortions. These anomalies, which the fuzzy logic of the human mind can more readily filter out and compensate for, can cause automated NMR interpretation algorithms to produce unreliable and extraneous results. Clearly, a new approach is needed.

Back in the early 2000s, Krish Krishnamurthy was leading the NMR Group at Eli Lilly and grappling with this very problem. He was frustrated by how long it was taking to analyze NMR experiments and obtaining good quality interpretation required the limited resources of the senior NMR experts. Krish and his team came up with the idea to employ the Bayesian approach (developed by Larry Bretthorst at Washington University), which meant not looking at the spectrum at all, but taking the free-induction decay (FID), time-domain data and analysing what comes out of it in a completely objective way. And that was the beginning of CRAFT (Complete Reduction to Amplitude Frequency Table).

CRAFT uses a Bayesian statistical approach to convert NMR time-domain data directly to the tabular domain. This bypasses the artifacts created by the Fourier transform when producing the spectrum and makes the automated extraction of these chemical shift and intensity data tables more reliable. With CRAFT, the spectrum (or frequency-domain data) is simply a human visualization tool for the data tables. This is in contrast to ‘conventional processes’ where the tabular domain is created from the manual interpretation of the spectrum by the spectroscopist. We have been collaborating with Krish to bring this new and revolutionary approach to our NMR customers. We recently sat down with him to discuss its evolution and potential.

We asked Krish what is special about CRAFT:
In the conventional way of doing analysis, the source of the information is the spectrum, which is the frequency domain data. Frequency domain data is two dimensional from an information point of view. Taking it into the tabular domain makes it amenable to automated and objective processing, as we are now dealing with numbers rather than a picture. The frequency domain spectrum is a picture, no more and no less, and that is the fundamental difference between analysis by CRAFT and analysis by conventional methods. In other words, A spectrum isn't a source of the information, rather a representation of the information.

CRAFT is redefining how NMR data is looked at by spectroscopists. It has redefined some of the fundamental practices in NMR because they were all based on frequency domain. A classic example of how CRAFT is benefitting users in unexpected ways is in the field of antibodies. Analysts tend to do a CPMG type of experiment to suppress the broad background signals. This background anomaly in frequency domain comes because you’re doing a Fourier Transform. When you think about it, it doesn’t make sense to add complexity to your experiment in order to fix a problem you have created by the way you choose to do your data processing. Just change your data processing method to not introduce this artifact. Similarly, when it comes to phase correction, this is done to make the frequency domain, the picture, look better. But if the picture is not the source of the information, then you don’t need to do phase correction, because there is no error and nothing to correct. With CRAFT, we are showing that you no longer need to do these adjustments anymore.

Talking about the challenges in promoting CRAFT to NMR users, Krish commented:
We are fighting the conventional wisdom of the user. The conventional wisdom for the last 50 years has been that the spectrum is the source of information, but really it is not. We are not taking away all the knowledge we have gained in the last 50-60 years of NMR evolution, we're asking people to look at it in different way using the same knowledge. We are slowly getting that message through to the market, but it will take time for CRAFT to become the accepted norm. However, we are confident that one day, it will be.

We asked Krish what he thought held the most potential for CRAFT in the future:
There are two main areas of significant potential for CRAFT, the first one is working on in-vivo imaging spectroscopy. There is significant potential to make decisions much more objectively because the materials are biologically divergent. There is an underlying increase in uncertainty and error by definition of these projects. By introducing an analytical technique that is less subjective, that increases their value.

The second area is pattern recognition. One of the things that we are guilty of in NMR is using the zoom tool to expand the region and look inside the spectrum, and we need to get out of doing this. When you think about it, structure elucidation of organic molecules is a simpler problem than many. We are working with a handful of elements, primarily C, H, N, O and maybe a couple of others. There are well-defined and simple rules about how these can be connected. NMR gives us clues to narrow the structure possibilities further, via fairly predictable and well-documented ways. When you compare that with the many complex things the world is attempting to create algorithms for, such as facial recognition, weather forecasting, self-driving cars, etc., there must be ways NMR can leverage these pattern recognition algorithms. But first we must get the data into computer-readable formats, reliably, reproducibly, and without artifacts.

Talking about the collaboration with JEOL, Krish commented:
The tool was originally built based upon what one person thought the user would do, through the collaboration with JEOL, they brought in many real-life applications, which helped us to adapt and evolve the approach to extend its usefulness. The collaboration asked quite a few practical questions about the way the tool was built and designed. This challenged us to look at how and why we were doing things in certain ways. It was really very valuable. CRAFT is an evolving technology, we are taking a staggered approach, and continuing to develop it thanks to dynamic feedback from NMR specialists such as JEOL and its customers.

To find out more about how CRAFT is transforming NMR experiments, download our app note describing how it was used for quantitative mixture analysis of pharmaceutical compounds: Download the App Note.

JEOL offers CRAFT for Delta V1.0 with Delta V5.3.0 software. Integration of CRAFT allows the JEOL NMR user to automatically and efficiently extract the best amplitudes and frequencies from NMR data.

On the NMR Support web site, the following resources are available for registered users of Delta and how to use CRAFT with the Delta NMR software:

Comments are closed.