SIMCA or Pirouette for OPLS-DA Log Out | Topics | Search
Moderators | Register | Edit Profile

NIR Discussion Forum » Bruce Campbell's List » General, All others » SIMCA or Pirouette for OPLS-DA « Previous Next »

Author Message
Top of pagePrevious messageNext messageBottom of page Link to this message

Bernard North (bnorth)
New member
Username: bnorth

Post Number: 5
Registered: 5-2007
Posted on Thursday, March 19, 2009 - 11:17 am:   

many thanks all for your advice .. v v helpful !

I understand now: the loadings for the dominant component (ie the non OSC one) in Pirouette wil be VIP like, thanks Scott

And its great that SIMCA 12 does do CVed scores plots - I've found them : go to Plot/List, choose scatterplots and select tcv for the x axis ! and again the VIPs should be like the loadings for the predictive score (except VIPs are positive), thanks Lennart

and I will certainly look into QDA and SVM thanks Venkatarman
Top of pagePrevious messageNext messageBottom of page Link to this message

Ian Michael (admin)
Board Administrator
Username: admin

Post Number: 18
Registered: 1-2006
Posted on Thursday, March 19, 2009 - 10:56 am:   

I have been asked to post this in reply to the original question. - Ian Michael

You raise several interesting and important questions, and I will try to cover them all.
In SIMCA-P+ version 12 it IS possible to plot cross-validated scores.
Regarding OSC and OPLS, the former was our first approach towards including the Y when constructing the filter of X. OSC is easy to overfit and care should be exercised when using this method. For this and other reasons we opted for another and more reliable solution, less prone to overfit, and so OPLS was developed. OPLS contains an "OSC-like" element, but the point is that the filter is included inside the projection model, not outside. Hence, OPLS is just ONE model, and we can avoid the two-step setup of using OSC + PLS.

The three-group problem is common in omics. A typical set-up is a Control group (or WildType group) vs. two separate treatments. Umetrics' approach to this problem would imply that two separate OPLS-DA models are developed, i.e., Control vs Treatment 1 and Control vs Treatment 2. The results of these models are then combined using the so called SUS-plot (see: Wiklund & Trygg, et al., Anal.Chem. 2008, 80, 115-122).

Furthermore, in the two-class discriminant problem OPLS guarantees ONE single Y-predictive component. This is not possible with the OSC-based methods. Any other systematic variation in the X-block will end up in one or more Y-orthogonal components of the OPLS model. The implication of a single Y-predictive component is easy discriminatory profile interpretation. For this latter purpose, we recommend using the so-called S-plot in connection with OPLS-DA (see: Wiklund & Trygg, et al., Anal.Chem. 2008, 80, 115-122).

It is possible to use the VIP-values of the predictive component for interpretation and these would resemble the corresponding loadings. However, one should be aware that VIP-values are always positive, while the loading (p) in OPLS-DA, which is central in the S-plot, has a direction so that an up-regulated metabolite gets a positive p-loading value. This latter fact simplifies interpretation from a biological point of view, since much attention is usually given to uncover which metabolite-groups are up- or down-regulated.

In addition, we also have the % explained variation diagnostics for both the Y-predictive and Y-orthogonal components. This information is helpful for the total understanding of the OPLS or OPLS-DA model.

All the best

Lennart Eriksson
MKS Umetrics AB
Top of pagePrevious messageNext messageBottom of page Link to this message

venkatarman (venkynir)
Senior Member
Username: venkynir

Post Number: 80
Registered: 3-2004
Posted on Thursday, March 19, 2009 - 1:01 am:   

Dear Bernard,
Your question is nice and even though I don't have much knowledge in chemistry , I can suggest the method building.
1. Try with QDA ,Mahalano
2. SIMCA is good but you should have good more number of samples for it
3. What about SVM , it proability based and new area for DA
think off
Top of pagePrevious messageNext messageBottom of page Link to this message

Scott Ramos (lsramos)
New member
Username: lsramos

Post Number: 5
Registered: 1-2007
Posted on Wednesday, March 18, 2009 - 12:39 pm:   

Bernard,

Pirouette avoids OPLS because of the patent, instead using a direct OSC, but the results should be similar whether you use the product from Umetrics or Infometrix. Because the VIPs are derived from the loadings, scaled by the explained Y variance, their use will be similar to the dominant loadings as well as the regression vector.

Regards,
Scott
Top of pagePrevious messageNext messageBottom of page Link to this message

Bernard North (bnorth)
New member
Username: bnorth

Post Number: 4
Registered: 5-2007
Posted on Wednesday, March 18, 2009 - 6:26 am:   

Dear All,

I wonder if I can have advice on which package to use for OPLS-DA.
I wish to discriminate 3 groups based on metabolites and to identify which metabolites are responsible for the separation.
I wonder whether Umetrics SIMCA or Infometrix Pirouette is best.
I'd like cross-validated scores plots - does SIMCA 12 do these ? The manual doesn't mention it.
I'd also like VIPs to identify which metabolites discriminate. Does Pirouette produce these ?
I think loadings only identify which metabolites contribute to the scores so may only explain X rather than predict class.
But if I do OPLS-DA then presumably the loadings for the discriminatory component(s) might be similar to VIPs ??
In which case I can dispense with VIPs and mainly use OSC filtered PLS-DA in Pirouette and use the first component loadings for the first cross-validates score.

I may use both package but often when I fit the same thing (eg a PLS-DA with 3 components) the results from the 2 packages aren't quite the same eg the scores and loadings are different, though I'm sure they are equivalent.

sorry this is a long question - many thanks in advance for any advice

Add Your Message Here
Posting is currently disabled in this topic. Contact your discussion moderator for more information.