I don't know a specific program, but as for a suitable algorithm, one could use the same approach that is used in image registration. For two chromatograms, one of them would have the retention time transformed, perhaps linearly (tnew = a*t + b) and your algorithm would simply look for the combinations of a and b to give the best registration of one chromatogram with the other (transformed; so this becomes a global optimum search on a plane). The best registration would simply be the highest similarity of the two spectra. As you can see, it boils down to finding the best way to compare the similarity of two spectra. The simplest measure is shared peak count, but one can use things like cosine distance or mutual information.

Thank you very much Matej,

i think the transformation of retention times is quite a solid approach. I am quite new to programming and never coded an optimization problem. Could you give me a hint how to get the optimal parameter values for the linear transformation here?

Martin, first of all I would search for special purposed software. It could be easier than coding it yourself, maybe someone else here will recommend a specific program. For example, I know two papers that describe protein spectra registration: http://arxiv.org/pdf/1311.2105.pdf and http://membres-timc.imag.fr/Sophie.Lambert/papier/Spectrometry.pdf and the software SpecAlign http://bioinformatics.oxfordjournals.org/content/21/9/2088.long It is for mass spectrometry, but may work for other kinds of spectra as well.

As for your specific question:

If there was a peak or a number of peaks that were always present in both samples, an initial estimate for a and b could be obtained by linear regression. Otherwise they could be limited by reasonable ranges. Say if the drift was not expected to be more than 10%, parameter a would fall between 0.9-1.1 and b between 0.9*tmax - 1.1*tmax. The step by which these values could then be changed in a numerical optimization depends on typical peak width. If it is, let's say 0.1% of the chromatogram width, then a and b could be incremented or decremented by a number slightly smaller than 1/1000th (or 1/1000th of tmax for parameter b). In a brute force approach this would mean a little bit more than 100*100 = 10000 chromatogram similarity calculations for every combination of a and b, to find the best combination of a and b among them. I am quite sure this is not the most efficient apprach, though. However, because the peaks are discrete, one probably cannot do things like gradient search to get the optimum.

As a fallback solution, you could also just save the chromatograms as images and try to do registration with appropriate image registration software, probably using more sofisticated methods than the one above, e.g. like the MATLAB code here: http://www.softpedia.com/get/Science-CAD/Image-Registration-Technique-for-Recovering-Rotation-Scale-and-Translation.shtml or the code here http://www.doc.ic.ac.uk/~dr/software/

Of course, the image-based approach will necessarily be inefficient, because the spectra are really just curves, one-dimensinal objects and we are forcing them into a two-dimensional analysis. Best results should be obtained with solving the "non-rigid curve registration problem", as is the case in some of the papers in the first paragraph of this message.

Not sure if is really a fit for what you are looking for, but you might give MeltDB a try:

http://meltdb.cebitec.uni-bielefeld.de

http://www.ncbi.nlm.nih.gov/pubmed/18765459

And as you are from Bielefeld, either my colleague or the people from the BRF might be able to help you with your problem.

TargetSearch--a Bioconductor package for the efficient preprocessing of GC-MS metabolite profiling data, it was developed on MPI in Golm, where metabolite profiling was introduced! so it works well done and is validated. on the over hand there a lots of different methods software now, sogar in Bielefeld entwickelt..

Grüße

Christian

thanks to all! Im gonna look through the suggestions now and Im sure something will fit.

I've aligned many, many chromatograms, and I always do them manually. If they are complex and you're working with samples that might have novel compounds, it's important to take the mass spectra into account when you are aligning (most automated alignment programs, as far as I'm aware, only take the retention time into account).

@ jeremy: Manual alignment is what was suggested to me too. Actually my samples are scent samples from elephant seals and were collected in the field. Many samples have low concentrations which often leads to noisy and broad peaks instead of sharp ones and even the mass spectra may be misleading here, or ? And don´t you think that manual alignment (although it may be more exact in the end) reduces reproducibility and may cause (expectation) biases?

Martin, that sounds very interesting.. Would love to hear more about it. I've worked with a lot of field-collected samples that were collected >10 years ago, and as long as they are stored properly it should be fine (quite a lot of work from a number of species has looked at how storage affects glandular secretions). The most important thing, in my opinion, is that all of the samples are handled in a similar fashion, which includes making sure that they are run on the GCMS around the same time and on the same column. Broad peaks can be caused by different things, and may or may not be resolved by altering your run parameters. This advice may be a little late coming since you already have your chromatograms, but see Drea et al (2013, American Journal of Primatology) for more suggestions on GCMS methods for biological samples.

To answer your question, I would say that I am more willing to deal with the possibility of human bias than with the possibility of calling one compound something else because the peaks have similar retention times despite different mass spectra. If you're really worried about expectation bias, you can have different people (at least two) score a subset of your chromatograms independently and then calculate interobserver reliability, either for entire chromatograms, particular sections, or perhaps for individual peaks. This approach is very common in studies in which people use ethograms to score animal behavior videos. In any case, if you do decide to just rely on retention times, you must be sure to note that you are disregarding your MS data; that is, you are only doing GC, not GCMS.

I would totally agree with you. Our approach is now somehow in the opposite direction. We also sampled DNA from all the seals and did a relatedness and heteozygosity estimation with microsatellites. It seems that from all 200 elements just a few are necessary to detect mother offspring similarity. Of course, a manual check whether the important substances are aligned correctly, also comparing mass spectra seems quite necessary. So you are working on lemurs? Always good to exchange with people, especially as scent analysis with natural populations is really underinvestigated!

Yes, I've worked on lemurs, other primates, and hyenas. Very interesting stuff you're doing. In addition to looking for correlations between genetic and chemical distance (for relatedness), you might be interested to look for correlations between chemical complexity (as measured by richness, simpson's index, or shannon's index) and genetic complexity (ie, heterozygosity). Total richness would certainly be the easiest measure here, because it requires the least amount of aligning (just take a total count of all compounds, after ensuring of course that you aren't eluting junk). Happy to chat with you about any of this or recommend some papers. Cheers!

Hi, I'm getting in a little late here, but I agree with Jeremy - I aligned chromatogram peaks manually. I was working on scent samples collected from wild echidnas; the results of GC-MS analyses were complex and we identified several homologous series' of different compounds (many of which were not in our search databases). If we'd left it up to an automated program, many of the peaks would have been incorrectly assigned. We also had a few problems with contaminants, particularly for poor-quality samples which had low detections for 'real' compounds - in this instance I used targeted mass spectra searches with a minimum peak size and mass spectra quality threshold to look for particular compounds. I kept human bias to a minimum by handling and preparing all samples in the same way throughout. I also re-did the quantification part for some early samples to check that my method was consistent across the study. I'm keen to hear more about your work on elephant seals - which population are you working on?

Hi Rachel,

you are definitly not late, I´m still in the project. It would be interesting to hear about your findings concerning scent profiles in echidnas. Did you do some genetics too?

With the elephant seals I seemed actually a bit confused, that was another project of mine. The samples are from Antarctic fur seals from South Georgia Island where the British Antarctic Survey has a quite cool research station.

Cheers, Martin

Hi Martin,

More details on the GC-MS analyses of echidna secretions are in Harris et al. 2012 Chemical Senses and Harris et al. in press J. Zoology (I can send you copies if you need).

I imagine elephant seals and fur seals would be quite challenging to work with? Is the project for a PhD/Masters, or something else?

I wanted to use genetic analyses too, we used microsatellite markers to see if measures of genetic quality and relatedness were correlated with chemical complexity or distance, so potentially function in mate choice (similar to Leclaire et al. 2012 Proc. Roc. Soc. B and Boulet et al. 2010 J. Exp. Biol). Unfortunately that wasn't successful.

Cheers,

Rachel

Dear Martin, I am way too late but here is a free alignment software MetAlign: http://www.wageningenur.nl/nl/Expertises-Dienstverlening/Onderzoeksinstituten/rikilt/show/MetAlign.htm

And why don't you visit with Prof Caroline Müller, Chemical Ecology, Bielefeld University. She may work with insects and plants, but certainly knows about peak alignment and PCA problems

XCMS work well if you want to use it for profiling and finding significant features between two or more groups 

https://xcmsonline.scripps.edu/index.php

Comes a bit late, be we wrote a little program in R to align GC data, which has especially been created to work with data from animal swabs, i.e. where there might be a lot of individual variation.

I attached a link to GCalignR, the vignette and the paper.

cheers

https://cran.r-project.org/web/packages/GCalignR/index.html

https://cran.r-project.org/web/packages/GCalignR/vignettes/GCalignR_step_by_step.html

http://www.biorxiv.org/content/early/2017/02/22/110494