GC-FID data: PLS, SVM, FDE ?

Question

Hello,

we are in contact with a company using GC-FID (Gas Chromatography-Flame ionization detector) as analytical tool.

They are comparing Raw Materials and want to understand which Analytical property leads to better performances.

Does anyone have this kind of data ?

PLS seems an appropriate tool but I've seen SVM in some publications, any thoughts ?

I was thinking also about FDE but is it worth the extra cost if PLS is the traditional tool ?

Thank you.

Emmanuel

@Phil_Kay @Bill_Worley

Daniel_Valente · Accepted Answer

Do you happen to know if the data are in the frequency domain or time domain?

ETA: looks like frequency domain. The Gasoline NIR data set would likely be close. PLS is used here.

Assuming something like this: https://www.researchgate.net/figure/Gas-chromatography-flame-ionization-detector-FID-results-of-the-outer-nest-tube_fig1_43285974

Peter_Hersh · Accepted Answer

@Emmanuel_Romeu ,

PLS has long been the industry standard for this type of analysis.  I personally think FDE works wonderfully for this type of data especially when you factor in all of the data manipulation options like DTW.  Analyzing chromatography data can be a challenge as the runs can be different time lengths (at least in R&D setting) and that is very painful to analyze with PLS.  @HadleyMyers has a great example of the pains of using PLS when the data is not in perfect form.  Here is the slide the basic message being analyzing this with PLS took months and resulted in a research paper.  This can be done in minutes with FDE.

Bill_Worley · Accepted Answer

Emmanuel,

FDE is a nice way to compare chromatograms, but is not necessarily a tool you would use for quantifying that kind of data.  However, if you have a target curve it can be very useful.

GC chromatograms are based on elution time of a given vaporized compound after it interacts (or not) with a stationary phase in a chromatographic column.   All chromatography is based on similar principles in general.   The mode of detection can vary greatly.

After a sample is injected into the oven it is flash vaporized and this mish-mash of gases is carried onto the column by a carrier gas.  Usually helium, but hydrogen is also used as carrier gas.  GC can be used to identify or quantify.

Identification is based on injecting known compounds into the chromatographic system and determining the retention time as they elute through the detector, FID in this case.  When the compound hits the detector it "burns" and generates a signal that is recorded in a "chromatogram".   This may be a reasonable use for FDE with a target curve.

Quantifying is more involved and requires preparation of solutions containing known amounts of the compound(s) of interest.  You would develop a calibration curve from this type of analysis.

PLS is reasonable if the whole chromatogram is recorded and they have timepoints as column headers similar to the Near-IR spectroscopic data in the gasoline data that @Daniel_Valente recommended.  Every row is one sample.

This data is very wide and that is why PLS works so well.  I have found Generalized Regression also works very well using a scanning Elastic Net with spectral data  and Neural Nets are worthwhile to try as well because of their flexibility.

I would suggest contacting @JeremyAshJMP as he has been working with chromatographic data and a peak finding algorithm in the FDE platform.

Sorry for the lesson on chromatography, but I hope that it helps better understand what your customer is trying to achieve.

Best,

Bill

Mark_Bailey · Accepted Answer

You might be able to quantitate results if the FPCs were calibrated. A simple example, prepare samples with know composition, obtain the functions, then use the composition variables as supplementary information. Like a FDOE.

Phil_Kay · Accepted Answer

Lots of good replies here.

I would just add that you should first check on some of the details of  their data analysis challenge. First of all, what form of data do they have: raw chromatogram or summarised data in the form of a "peak table"?

With chromatography methods the instrument software will often identify the different peaks in the chromatogram for each sample and produce summary statistics (integrated area, width, symmetry, ...) for each peak. Often a data table of integrated area for each peak for each sample is what they need to solve their scientific problem. Basically each different chemical entity in a sample will produce a peak at a unique point on the time/x-axis of the chromatogram. The area of the peak is proportional to the amount of that entity in the sample.

Let me know if this is what they need to do. I have an example that you might be able to use but I don't want to share it widely.

If they need to use raw chromatograms then FDE could be useful, as others have mentioned. There are some examples of customers doing this already in certain circumstances. But FDE capabilities still need considerable development to be really useful in the majority of chromatography examples.  Currently FDE will only be useful with very simple chromatograms and it doesn't sound like this will be the case in your example. I don't know of any example data sets that you could use.

I hope this helps.

Phil

HadleyMyers · Accepted Answer

Hi Emmanuel,

I think this is the dataset Pete is referring to.  It's simply the fermentation data converted to wide-format.  Anyone who tries to analyze it using PLS will find that it's longer and much more complicated than the FDE workflow, and still doesn't provide answers that are as easily communicated.