Chemometrics - Science topic

You can find free datasets related to gravimetric and spectral methods on websites like Eigenvector Research, RamanSPy, Mendeley Data, and Nature, which offer various spectroscopic data for analysis. These resources provide valuable datasets for research and algorithm testing in analytical chemistry.

Choosing the lowest RMSECV would lead to overfit the model and it is not good.

Before you perform PLSR, I suggest to understand the data first. Following steps may help you to analyze your data:

1. Analyzing its PCA to find some pattern or outlier identification

2. Analyzing the raw spectra to examine the most prominent peaks that has chemical information and can contribute to targeted y-reference. This helps to check with regression vector later to avoid an overfit model. Other thing, this may help to indicate, maybe, some strange spectra.

3. Try to see number of data. Is it possible to split the data as training and independent test set? rule of thumb to split is 2/3 for training and 1/3 for test. This is important as it allows you to see whether the model tends to overfit or not. Cross validation could be the last option if number of data is limited

4. In the case this research only can use cross validation, after building the model, see the revolution of RMSE. Most of model have a trend where the more LV used the lower RMSE will be but at certain point, it won't improve that significant, so, should stop on LV before that point.

5. Try to cross check with regression vector to see whether the model try include noise or only prominent peaks that responsible to targeted value. if the reg vector has less noise, it could be an indication that your model is robust.

Has your sample signal been normalized? The differences do not appear to be very pronounced. Additionally, why have you chosen to use Ti 441.492 nm? It is clear that this is not a commonly used spectral line for Ti. There are many strong peaks for Ti that could be used for univariate calibration, such as Ti I 365.35 nm, Ti I 375.29 nm, Ti I 399.86 nm, Ti II 323.45 nm, Ti II 334.94 nm, etc. Why have you not selected these? Are there other interfering peaks? Self-absorption occurs when the Ti concentration is too high, and in such cases, it is advisable not to use easily excited characteristic peaks as they are more prone to self-absorption. You can simply determine whether self-absorption is occurring by checking for peak center dips or splitting. If you need to quantify the degree of self-absorption, it can be calculated using the appropriate formula.

These figures show PCA scores projected onto their corresponding pixels.

Dear Joseph,

Greetings from SCIENCE4U India, hope you are doing well and the war is less influencing your research activities.

Let me know if you need any assistance for your research work from my side.

Best regards,

Bhanu.

"untargeted analysis" is fine, if the data are valid, but they are not in this case and using them as you suggest is invalid. In any case, now Mohamed Mahmoud has been provided with professional advice that he should not use the raw area % units from his analysis data outside of the one analysis method. To write any more would be repetitive and a waste of Mohamed Mahmoud's time as his question has been answered.

I wish you luck with your classes at school Jokin and thank Mohamed Mahmoud for his patience.

Dear Sarang Jokhio

I am a bit late to answer your question but let me try to explain to you as simply as I can.

The continuous dependent variable (y) is used to develop a predictive model. Of course and as you did, you need a multivariate matrix Xnxp, where n is the number of samples (rows) and p are the features (columns). Then, with PLS regression you obtain the weights (regression coefficients) that maximize the covariance between X and y. With these regression coefficients, you are able to predict any future or different dataset.

In terms of the categorical variable (y) is used to develop a classification model. Here you need the matrix X as well but need to use different algorithms such as logistic regression, as David Eugene Booth explained, or PLS-DA, LDA. The goal is to obtain the weights or coefficients that minimized the error. Those, as in regression, with those coefficients you will be able to classify any future or different dataset.

Now, I will explain your figure. Since you are dealing with a regression problem and you used PLS regression the first step is to optimize the PLS number or component or latent variables. This step could be done using the left-bottom plot. As you could see, it is related to RMSE, which is an error, therefore lower it is the better. This figure showed that as the components (x-axis (PLS components or latent variable)) increased the error RMSE decreased in both cv (red line and means cross-validation) and cal (the blue line which means calibration). I can see that you or the software suggested the number of components (n_comp =3) based on the other plots. But as you could see, at component 3 both curves have not reached the lower error, in my case, I will select the number of components 7 or 8 since both curves have reached almost the minimum and they look close to each other meaning that the model is generalizing better compare for example with component 5 or 6, in those components, the red curve suffers an increment.

Then, the remaining plot will be based on the selected PLS component and for sure component 7 will give you a better predictions plot (right-bottom). The top-left plot could be used to select outlier samples and the top-right plot is the regression coefficients that as I told you above could be used for predicting or classifying future datasets.

I hope that my explanation makes sense to you and could be useful. Do not hesitate to ask if you have any other questions,

All the best

Something like XLStat can be run directly from Excel. Please have a read here if it helps: https://high-tech-guide.com/article/can-you-do-principal-component-analysis-in-excel

If you want something free, you might look into Python packages, but maybe too advanced.

Dear, Amani Albalawi

I suggest that the easiest and the cheapest way is using web-version of MetaboAnalyst (https://www.metaboanalyst.ca/MetaboAnalyst/ModuleView.xhtml).

Best regards,

Ivan

Dear Joel Johnson

you can contact me and Dr Silvio David Rodriguez for possible collaborations.

Regards,

When original data is modified, then information content is lost. The missing information may or may not be critical to the parameter estimates of interest.

Suppose one data set was modified and you have a series of unmodified data sets, now objective prior information is available and it might be possible to develop a statistical model for the lost information in the modified data set. For example, if one data set was modified to remove a constant (such as a DC baseline offset). You could create a statistic to model the DC offset using the unmodified data sets. Now the estimates for the original data is based on objective prior information.

If someone alters the data and you have no idea whatsoever about the changes they made, then there is no hope. Any method that attempts to recreate the original data is subjective.

In your example - " assumes(s) your empty channel spectrum changes within the series), large enough so that it matters, but small enough that the data still seem to make sense" - this could be objective prior information. If the model for the data could be modified to represent the missing data, then Bayesian analysis would account for the missing data (empty channel) using objective prior probability distributions for all the parameters in the model for the data. If the resulting parameter estimates of interest turn out to have very large uncertainties, at least you would know the unmodified data was inherently uninformative.

You can also use free online tools for multivariate statistics: MetaboAnalyst (https://www.metaboanalyst.ca/) or Biostatflow (http://biostatflow.org).

p.

Hi Duri,

I organized some information about PCA and PCR on:

The original is in Portuguese but can be translated by Google Translator:

Best Regards,

Markos

I suppose that non-volatile compounds are unable to vaporize and thereby, cannot undergo sublimation, and can not be detected by electronic nose.

Another possibilities include decoupling your main matrix problem (after preprocessing it) into some combination of matrices with known structures (e.g., Toeplitz, circulant, band diagonal, etc.) and, then solve the subproblems according to your needs. Tensor structures/operations may be optimized for some structures in terms of hardware and networking.

As an engineer, I like to think that the problem and model tend to give some clues regarding simplifications.

The preprocessing stage depends on what you are investigating. Sometimes, outliers have more info than mainstream data.

Finally, the word "compression" has different meanings, levels and varieties. If you change your domain for a given tensor structure, you may get some constrains/compression.

Thanks for sharing

To Dr Alejandro C. Olivieri,

we are interested in using the LOD calculation as described in your publication :

Franco Allegrini and Alejandro C. Olivieri (2014)

IUPAC-Consistent Approach to the Limit of detection in partial least squares calibration.

Analytical Chemistry, 86, 7858-7866.

We have some difficulties in applying equations (12) and (13).

1) Does the term var(ycal) refers to the variance of the whole set of y calibration values (that is the n values associated with the n spectra forming the calibration set) or  to an evaluation of the error of the measurement of y ?

2) ycal is said "centered" in the publication. In this condition, the mean of ycal  is compulsorily equal to 0 and cannot appear as a denominator in equation (10).

2) As a spectrum is a vector, what does mean var(x) ? We suppose that it is the mean of the variances as measured to each element of the vector x.

Do you have a simple numerical example on which we can compute the LODs ?

Thank you in advance for your valuable help,

Kind regards,

Dr Dominique Bertrand

--

Dominique Bertrand

chimiométrie et data processing

0633338680

Dear Sindhuraj Mukherjee,

I suggest you to see links and attached files on topic.

Best regards

Yes, the folder looks like this for a single snap!

I will inbox you as well. Thank you again :)

Muhammad Ali Thankyou for the article recommendation.

Julien Lamour

Apologies - in the example here - the Loadings are the files of interest

https://uk.mathworks.com/help/stats/partial-least-squares-regression-and-principal-components-regression.html;jsessionid=6be9925bb119234705652e1765a7

Yes. I use the frequency selected trough algorithm of sucessive projection by vectors in n - euclidian space, where the variables projected in space one - by - one, if equal zero is excluded.

The beer law can be aplied because the linear equation form. But, is needed select variables ( frequency ), trough test with number minimum and maximum of variables indiviually.

Kirk,

I want to use PLS or some other multivariate chemometric technique for their ability of using tons of variables for the predictive model. In the case of FTIR, generally the integration of one spectral region is used in a traditional concentration curve. In the case of PLS, I can use the entire spectrum (or a good portion of it) and selective the important LV.

Probably for you variables are the components in the mixtures. In my case, there are generally between 5-10 components in the mixture. I am interested in the quantification one of them only. The other 4-9 components create(correlate) with the signal of my target in the entire spectrum. The issue with synthetic mixtures is the poor quality of the standard materials (with the exception of my target) and limitation of the techniques to determine the composition of the standard materials. The issue of understanding exactly what is in the mixture is why I cannot use simply real samples.

The standard addition method could be the best way ....but....I have not figured out how to do it chemometrically.

Does it make it sense?

Principal Component Analysis (PCA) is normally able to gather similar data and discriminate between different sets. We successfully did it with coffee aroma analysis using GC-MS.

PLS-DA is a supervised pattern recognition technique, used to classify the unknown samples to the predefined class.

First you need to

In this technique, each sample in the calibration set is assigned a dummy variable as a reference value (you need to add new column and assign each sample to its class), which is an arbitrary number designating whether the sample belongs to a particular group or not for example (1 = Class A and 2 = Class B). A sample is classified as A if its value was below 1.5 and classified as B if the value was above 1.5.

The following paper applied the PLS-DA technique for classifying honey based its botanical origin.

Check this guys in barcelona they are doing good:

Good question

yes it can be

For curve fitting, a mixture of Gaussian ans Lorentzian is most appropriate. As suggested by Adeyinka Aina PCA can be useful for comparing 2 samples. PLSR is usual for continuous changes over multiple samples, but assumes a linear correlation to an external variable (e.g. concentration, temperature). MCR-ALS can by used for more general correlations.

Dear Giulia,

Near infrared (NIR) and Raman spectroscopies have already demonstrated their potential for in-situ analysis. It is nevertheless difficult to say in advance which one would be the best for this specific application. it must therefore be tested.

Best regards,

Ludovic.

Hello, first of all, what's the sample amount in each group? In general, I suggest to test non-parametric if n < 6. For smaller sample size it is not really serious to assume parametric conditions, independently from test results. As non-parametric alternative for t-test, I suggest u-test (Mann-Whitney-Test). Best regards, Max

Dear Bhupendra Shrestha

Do you have the chromatograms for each of the individual API's? Is your detector a fixed WL or DAD? If you share (in a general format - Excel, tab delimites, comma separated values, etc) the mixture chromatogram data as well as the individual API chromatograms I can try to take a look and give you some feedback. Regards, Luis

Dear Praveen Kumar Yadav ,

I will give you some useful information related to PCA.

Scores: describe the properties of the samples and are usually shown as a map of one PC plotted against another.

Samples with close scores along the same PC are similar (they have close values for the corresponding variables). Conversely, samples for which the scores differ greatly are quite different from each other with respect to those variables. The relative importance of each principal component is expressed in terms of how much variance of the original data it describes.

Loadings: describe the relationships between variables and may be plotted as a line (commonly used in spectral data interpretation) or a map (commonly used in process or sensory data analysis).

For each PC, look for variables with high loadings (i.e. close to +1 or –1); this indicates that the loading is interpretable.

To study variable correlations, one studies the relative location of variables in the loadings space. Variables that lie close together are highly correlated. For instance, if two variables have high loadings along the same PC, it means that their angle is small, which in turn means that the two variables are highly correlated. If both loadings have the same sign, the correlation is positive (when one variable increases, so does the other). If the loadings have opposite signs, the correlation is negative (when one variable increases, the other decreases).

But remember: Loadings cannot be interpreted without Scores, and vice versa.

Loadings. For that reason the BI-PLOT is the best plot for analyzing PCA.

About the significance of each score and loadings, it depends of the software that you have used for processing your data. In the result, you should be able to find the P values of each score and/or loadings,

Wish you all the best,

You can use a GA-PLSR matlab code by Riccardo Leardi to reduce descriptors and then select the best descriptors by stepwise regression in SPSS software.

GA-PLSR code:

Application of genetic algorithm–PLS for feature selection in spectral data sets

https://doi.org/10.1002/1099-128X(200009/12)14:5/6<643::AID-CEM621>3.0.CO;2-E

Dear Niazi,

The molecular design will be more inclined to embrace a transcriptomics-based approach that centered on RNA-RNA, RNA-DNA, and RNA-Protein interaction studies. The methods to predict 2D and 3D structures of RNA are already determined as shown in this reference:

Parikesit, A. A. (2018). The Construction of Two and Three Dimensional Molecular Models for the miR-31 and Its Silencer as the Triple Negative Breast Cancer Biomarkers. OnLine Journal of Biological Sciences, 18(4), 424–431. https://doi.org/10.3844/ojbsci.2018.424.431

Parikesit, A. A., Utomo, D. H., & Karimah, N. (2018). Determination of secondary and tertiary structures of cervical cancer lncRNA diagnostic and siRNA therapeutic biomarkers. Indonesian Journal of Biotechnology, 23(1), 1. https://doi.org/10.22146/ijbiotech.28508

In this respect, a design of silencing (si)RNA molecules as drug candidates will gain more importance in the future, as well as studies of long(ln)RNA biomarkers. Keep in mind that although transcriptomics-based approach will gain importance, the proteomics-based one will still be utilized by molecular designers.

For UV-Vis or many other type of specroscopic titration data you can use HypSpec software. Check through the following website. The best I have ever known.

There’s a good chance that you could use Raman Spectroscopy to do this analysis. Because Raman has very fine peaks and it is qualitative, it is a very good technique to determine an analyte within a mixture. You would need to obtain the Raman spectra of the active ingredient and the expedients in the mixture, and determine a suitable peak or peaks to measure. PLS is a good Chemometrics technique to use.

Colleagues of mine have done similar work – check their article on ResearchGate:

Different related software's can be used for building ANN predictive model using NIR spectra. My suggestion are Unscrambler and IBM modeler. Before ANN modeling, PCA should be done to reduce large spectra variables to PC1, PC2 .... . after calibrating ANN, you have predicted values by calibrated model and you have also reference values for each sample. Now use RPD index to realizing the goodness of your calibrated model.

If this is what you mean (see attached images), I can recommend OpenChrom: https://www.openchrom.net/ You can shift the chromatograms left/right by clicking on arrow buttons.

BioRender! - (https://biorender.io/)

Is a free online app that contains a library of pre-made cells, proteins, membrane shapes, organs, lab equipment, etc. that you can drag-and-drop so you don't have to spend time drawing each element of the figure out yourself. Saves a lot of time for creating schematic figures, and the icons are all created by scientific illustrators so they're both beautiful and accurate! And it's free for educational use.

The best choice in such a confusion is either to check the calibration of the spectrometer or record the spectra from pure copper and iron sample and then compare it with your sample. The best is to record the spectra from pure samples and compare it with your data.

You need to model multi-variate regression (PLSR/PCR) between each laboratory obtained soil nutrient data of the sample points, keeping as dependent variable (Y) and spectral reflectance of the sampling points obtained from HYPERION data as independent variable (X). One usually predicts the nutrient value from the relationship between X and Y and cross-validates the predicted value through examining the precision of fit in linear regression between the actual and predicted values of nutrient.

You can learn about the MATLAB implementation of PLSR/PCR from the following link:

First you have to do is to assign observed bands to normal modes of the material you measured. To do that you need the structure of the material / compound which was measured.

Assignment of normal modes you can do either on the basis of literature or with help of quantum-chemical calculations. For calculations you can use Gaussian software package. Other options (packages) are also possible. How to use quantum calculations for assignment you can find under link of Spectroscopy and Molecular Modeling Group: https://smmg.pl/home-en-gb.

When your bands are assigned to normal modes you can compare two spectra: to look for the same and different modes (frequencies are important) and the contribution of modes (absorbance).

Software for spectroscopy: Grams or Opus (both are expensive). I do not know if you can find anything free of charge. Try to find the web-page Spectroscopy Ninja and negotiate free version of the software you can find there.

That is all what I can suggest generally. If I know how your spectra look like, conditions of the measurement and the structure of your materials (compounds which spectra you measured) I can try to make more precise suggestions.

Never used EPO but there is a Thesis Dissertation by Nanning Cao that can help

https://lib.dr.iastate.edu/cgi/viewcontent.cgi?article=4206&context=etd

After drying you most probably get a coffee ring effect, meaning that the thickness of you layers is strongly changing from point to point. Therefore imaging and measuring individual points might in any way not be the correct method for reasons explained in

The electric field standing wave effect can be corrected as explained in

My feeling is, however, that you should better use a different technique, maybe to coat an ATR crystal and see if you could obtain a calibration line, in particular also, because for near normal incidence and thin layers, you might see nothing at all on your metallic substrate due to the fact that a metallic surface suppresses electric fields parallel to its surface. Certainly, this still does not solve the problem of overlapping bands.

Hi,

I found this paper that maybe related with your research

”"Variety Identification of Single Rice Seed Using Hyperspectral Imaging Combined with Convolutional Neural Network"

There are also several papers that discuss about plant disease recognition using HSI and CNN.

"Explaining hyperspectral imaging based plant disease identification: 3D CNN and saliency maps "

Links: https://arxiv.org/pdf/1804.08831.pdf

"Classifying Wheat Hyperspectral Pixels of Healthy Heads and Fusarium Head Blight Disease Using a Deep Neural Network in the Wild Field "

Links: https://res.mdpi.com/def50200dc1e333db4fc797d476b6adff985faf79860636a7cc3541a61034885c872bbcc2ce188cdcd5a4a69b06c6186212c0e3e60365998e7ba6cd4bb76b0ca24da6b0201dcd947667f1fe5ae018c36ab81b4484681dd834ad9650a6970e256a06e83d5ca81801ac364d24ad6ef9c908ce5f71327cb89b949ad7cf73756ed6baa30053b3ff8bb1d25471d45ce194f2345f866cf7d1c08aa5fd5009f4c3fbac604826e1f65?filename=&attachment=1

Hope it helps your research.

Depending on your sample, you can go for supervised and unsupervised learning methods. pca and clustering techniques can be used for unsupervised learning whereas discriminant analysis, neural network models can be used for supervised learning

No, thank you.

Dear Sir. Concerning your issue about the use of FTIR and chemometrics as their SOP for microbial testing . IR spectroscopy is an excellent method for biological analyses. It enables the nonperturbative, label-free extraction of biochemical information and images toward diagnosis and the assessment of cell functionality. Although not strictly microscopy in the conventional sense, it allows the construction of images of tissue or cell architecture by the passing of spectral data through a variety of computational algorithms. Because such images are constructed from fingerprint spectra, the notion is that they can be an objective reflection of the underlying health status of the analyzed sample. One of the major difficulties in the field has been determining a consensus on spectral pre-processing and data analysis. This manuscript brings together as coauthors some of the leaders in this field to allow the standardization of methods and procedures for adapting a multistage approach to a methodology that can be applied to a variety of cell biological questions or used within a clinical setting for disease screening or diagnosis. We describe a protocol for collecting IR spectra and images from biological samples (e.g., fixed cytology and tissue sections, live cells or biofluids) that assesses the instrumental options available, appropriate sample preparation, different sampling modes as well as important advances in spectral data acquisition. After acquisition, data processing consists of a sequence of steps including quality control, spectral pre-processing, feature extraction and classification of the supervised or unsupervised type. A typical experiment can be completed and analyzed within hours. Example results are presented on the use of IR spectra combined with multivariate data processing. I think the following below links may help you in your analysis:

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4480339/pdf/nihms-700198.pdf

Thanks

Check out BioVinci. You just need to drag and drop to run PCA. Quite simple to use and easy to understand. You can go here to see the PCA plot example: https://vinci.bioturing.com/panel/workset/build/principal-component-analysis

Statistics and Chemometrics for Analytical Chemistry, a textbook by J. C. Miller, James N. Miller, and Jane Miller

Either can be reported, or both. If the transformation is simple and common, it sometimes makes sense to report the transformed values: "mean log10(concentration)". In other cases, you can report the "back-transformed" data, or data "on the original scale." In a plot, sometimes using the original data makes the plot unreadable.

Following

Hello,

Please look at the approach developed by Heien (e.g. Chem Commun (Camb). 2015 Feb 11; 51(12): 2235–2238. doi:  10.1039/c4cc06165a or google: "fast-scan controlled-adsorption voltammetry" ) and Blaha (DOI: 10.1021/acs.analchem.6b02605 Anal. Chem. 2016, 88, 10962−10970). Both methods require modification of Vappl. The FSCAV I tested and got positive results even without PCR. Note that in my opinion a solid pharmacology with a selective (in terms of DA/metabolites) compounds should be used by several labs before these methods may be considered as "established".

Regards,

Leonid

I finally rereredid the analysis with the script you gave me and I finally obtained something similar to the previous results obtained with MATLAB! so thank you to have encouraged me to try again ;-)

best regards

Dear Giorgio Luciano, can you suggest to me some real case study, or sample data which shows that the dimension of variance covariance matrix is not a square matrix when the number of variables exceed the number of samples when calculating Mahalanobis distance?

Hi Nicola, I know that multivariate analysis is the best approach in spectral identification but for my aim I have to compare two single spectra and verify how much they are close or different in order to distinguish them.

My principal idea is to use Pearson correlation coefficient or evaluate the std of the difference between the spectra. Anyway will be useful to know if more robust method can be applied (excluding the calculation of derivative). Some spectral matching tool like pattern matching for images?

Regards

Good day

You can download the reference blow.

Sincerely

Dear Sir. Concerning your issue about the Raman Spectrum Databases of Carotenoids such as Lycopene, B-carotene, and Lutein. I think the following below links may help you in your analysis:

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3086339/pdf/nihms277219.pdf

https://www.ncbi.nlm.nih.gov/pubmed/17165839

http://ajcn.nutrition.org/content/96/5/1179S.full.pdf+html

Thanks

Dear Dr. Tanamy

i need to cooperate with you to make analysis for three essential oils (already analysed by gc-ms and have antioxidant activity) by chemometric analysis as i tolled you before.

please tell me what the information you need about three essential oils

waiting your response soon

Prof. Dr. El-ghorab

Hi

Fist of all, why do you want to do a RSM (response surface design) ? being a central composite design, composite design in general, Box-Behken design?

- with this question I mean what do you want to achieve as a result from your experimental work:

A process/synthesis/unit operation optimization should be conducted by some steps

(1) Identify a procedure / method that operate

(2) Conduct a parameterization of the procedure, that is to identify the experimental variables that are possible to variety and the responses that can be measured (yield, selectivity, ….)

(3) Conduct an introductory screening design. In this you commonly identify the solvent, catalyst etc. that you should use. In your case you should consult. the paper by Rolf Carlson and collaborators: Strategies for solvent selection. Acta Chem Scand 1985 B39, 79-91 (and the book by Carlson see below).

When you have performed a screening design, where you should include as many variables as possible. Type of design for this purpose: I suggest you to use a fractional factorial design. By such a design you can identify which variables that are the more important and identify a preliminary optimum.

(4) The preliminary identified optimum might subsequently operate as the centre of a new design that allow you to create a predictive model that allow you to estimate: (a) coefficients for single variables, (b) coefficients for the square of each of the single variables and (c) the coefficients for two-factor interactions of the single variables. I strongly recommend to you to use a response surface design of the type central composite design. This allow you to estimate a model:

y(x₁, x₂, …, x_k) = b₀ + b₁x₁ + b₂x₂ + … + b_kx_k + b₁₂x₁x₂ + b₁₃x₁x₃ + … + b₁₁x₁² + b₂₂x₂² + … b_kkx_k²

Be aware that sometimes, the star points of the central composite design cannot be created; e.g. reaction time -1 (10 min)  0 (25 min) +1 (40 min) will result in that “-2” become a negative time:  -5 min.  In such cases you can solve this by change either the value of (-1) increase the reaction time, shorten the step, i.e. the (0) will be at a smaller value. You can also change the value of the star point.

Introduction of the categorical variables such as: “type of solvents”, “type of catalyst”, “type of reactor”, “type of base”,  etc. should not, from my point of view” be included in a response surface design. In fact, such variables shall be investigated by means of a screening design, where all such parameters (variables) is investigated/compared and decided before you do the response surface deign (a composite design).

Some “bibles” for statistical experimental design, screening and response surface modelling and more:

Box, G. E. P.; Hunter, J.; Hunter, W. G. Statistics for Experimenters: Design, Innovation, and Discovery, 2nd ed.; 390 Wiley: New York, 2005.

Box, G. E. P.; Draper, N. R. Empirical Model-Building and Response Surfaces; Wiley: New York, 1987.

Montgomery, D. C.; Peck, E. A. Introduction to Linear Regression Analysis; Wiley: New York, 1982.

Draper, N. R.; Smith, H. Applied Regression Analysis, 3rd ed.; Wiley: New York, 1998.

Rolf Carlson Johan Carlson

ISBN: 9780444515278, Elsevier Science, 8th April 2005

Hi

There is a good example of a study using Mahalanobis distance 

Rodriguez-Saona LE, Khambaty FM, Fry FS, Dubois J, Calvey EM. Detection and

identification of bacteria in a juice matrix with Fourier transform-near infrared

spectroscopy and multivariiate analysis. J Food Prot. 2004 Nov;67(11):2555-9.

PubMed PMID: 15553641.

but most groups  use Euclidean distance 

Cebi N, Yilmaz MT, Sagdic O. A rapid ATR-FTIR spectroscopic method for

detection of sibutramine adulteration in tea and coffee based on hierarchical

cluster and principal component analyses. Food Chem. 2017 Aug 15;229:517-526.

doi: 10.1016/j.foodchem.2017.02.072. Epub 2017 Feb 22. PubMed PMID: 28372210.

Zhang J, Huang P, Wang Z, Dong H. Application of FTIR spectroscopy for

traumatic axonal injury: a possible tool for estimating injury interval. Biosci

Rep. 2017 Jul 21;37(4). pii: BSR20170720. doi: 10.1042/BSR20170720. Print 2017

Aug 31. PubMed PMID: 28659494; PubMed Central PMCID: PMC5567294.

or squared Euclidean distance

Fuenffinger N, Arzhantsev S, Gryniewicz-Ruzicka C. Classification of

Ciprofloxacin Tablets Using Near-Infrared Spectroscopy and Chemometric Modeling.

Appl Spectrosc. 2017 Aug;71(8):1927-1937. doi: 10.1177/0003702817699624. Epub

2017 Apr 10. PubMed PMID: 28393531.

For general multivariate analysis of NIR data the following is quite helpful but doesn't address the distance matrix directly

Rosas JG, Blanco M, González JM, Alcalá M. Quality by design approach of a

pharmaceutical gel manufacturing process, part 2: near infrared monitoring of

composition and physical parameters. J Pharm Sci. 2011 Oct;100(10):4442-51. doi:

10.1002/jps.22607. Epub 2011 May 5. PubMed PMID: 21557224.

 best wishes,

Elaine

As  Edgar says MATLAB, Python and R are all softwares where programmes are available for spectral analysis. The research group who are world leaders in spectral processing including NIR and Raman are based in Umea university and have pretty much developed the field of Chemometrics. They use a commercial software package (SIMCA) as well as R/Python but their software algorithms are often published. If you take a look at the webpage of Johan Trygg who is the head of the department (http://www.chemistry.umu.se/english/research/group-leaders/johan-trygg) you will find lots of articles on how to process and model these types of spectra. They are an extremely friendly group and welcome interactions with scientists.

Hello,

I would also translate "proportion for the peak area of identifiable compounds based on NIST database to the total detected products used GC-MS" into ratio between "sum of area counts for identified peaks divided by area counts of the TIC" . This gives you some idea about how much (as % of the sum of amounts) of the already present compounds were identified. To my opinion this is not "meaningless" as Charles has stated, but i agree it is a limited tool. The area response of the several compounds are varying much. But EI ionization in GC-MS is a more universal ionization type compared to other instrumentation, therefore the comparison has its worth. You can support this additionally with GC-FID data where the area response of hydrocarbons is very similiar. It is easy to criticize this approach, but to my knowledge there is no other practicable chromatographic approach available to estimate the extent of the identification work in a sample. If this is note the case Wayne and Charles should provide us better alternatives.

By the way, I support Asit that apart from NIST matching and authentic standards, it is recommendable to calculate retention indices. NIST supports You with respective data.

 Hi Sandrine,

Have you finally found the information somewhere since? I found out that PET/chlorophenol alone should be too viscous for GPC and that a blend with chloroform (25% chlorophenol/75% chloroform) should be better [1]. I looked into the literature to find the dn/dc of PET/chlorophenol but only have found the PET/chloroform couple (=0.143 @ 25°C if you are interested in) in this book [2] (which can be useful to have for GPC in general):

[1] S.A. Jabarin, D.C. Balduff, Gel permeation chromatography of polyethylene terephthalate, journ. of liq. chrom., 5 (10), 1825-1845 (1982).

[2] A. Theisen, C. Johann, M.P. Deacon, and S.E. Harding, Refractive Increment Data-Book for Polymer and Biomolecular Scientists (Nottingham University Press, Nottingham, UK, 2000),

Hope it will help.

If you have found another solvent for your GPC measurements, I would be interested to know how (different from HFIP that I cannot use).

Thanks,

Violette

Thank you for this nice remark. We have investigated spectral differences directly in order to identify discriminating spectral features and their relation to biochemical differences. For chemometric modeling we are obtaining good results with Principal Component Analysis in combination with Linear Discriminant Analysis. Details have been described in Santos et al, Anal Chem. 2016, 88(15):7683-8.

TRY AQUA CHEM

Thanks Thomas, for the addition of the other normalization techniques.

To Liz: PQN is probabilistic quotient normalization; TotInT is area normalization, Range is similar to TotInt but instead of the entire spectra only a definied range is used; zero2one is also called unit vector normalization.

As described by Thomas you can calculate peak normalization in Excel or also in R. However, for peak normalization each intensity value of a single spectrum (i) is divided through the intensity of this particular spectra (i) at the choosen wavenumber (k):

X_i^´ = X_i / x_i,k

After this transformation you import data to chemspec.

However, it would be quite more comfortable to import the spectral data first to Chemspec doing baseline correction or binning (see workflow in figure 1, https://cran.r-project.org/web/packages/ChemoSpec/vignettes/ChemoSpec.pdf) and calculate peak normalization afterwards. But therefor you need to know how to get single intensity values of each spectra. Perhaps you can contact directly the package author and ask him.

This is based on the adsorption coefficients of the compounds to be identified. This is to answer your question in brief.

 Hi A. Abdo,

You typically need one ANN dedicated to multivariate function approximation. Radial Basis Function networks (RBFs) constitute such a good candidate solution to solve your issue (follow https://www.hindawi.com/journals/isrn/2012/324194/ ). They may be efficiently implemented as Convolutional Kernel Networks (CNNs, follow http://ceur-ws.org/Vol-1649/118.pdf ). From a practical point of view, there exist a MATLAB library supporting CNNs with an ad-hoc tutorial and code samples:  http://www.uow.edu.au/~phung/docs/cnn-matlab/cnn-matlab.pdf

As an alternative, Matlab Central provides with an implementation of RBFs, the initialization of which one is based on k-means ( https://fr.mathworks.com/matlabcentral/fileexchange/52580-radial-basis-function-neural-networks--with-parameter-selection-using-k-means-? ).

Further readings that may be helpfull to you about ANNs (see alternatively SVM networks) :

Regards

Check the like below

Step 1 : rename  .dpt to .txt

Open it in .txt and replace all comma (,) with 5 spaces (     ) and save it

Step 2: Now replace all . with , and save it

Step 3: Open it in origin.

Goodluck!

I have done this also.

But still I am stuck with those negative frequencies. Can you please help me out with this problem?

Thank you, Dr. Michael, for sharing the report. I believe, after the mass loss of HCN and CO, 1-amino-2-naphthol will be converted into Benzocyclobutadiene (which is a stable molecule).