In earlier Tutorials here at Chemometrics.se, I discussed why Partial Least Squares Projections to Latent Structures (PLS) has problems in dealing with strong and structured noise in the descriptor matrix X, hence causing the PLS model to include more components than Y-variables. Structured noise (Y-orthogonal variation) in X also causes problems for other projection based methods such as PCR and other methods with similar properties. I have also discussed a new set of pre-processing methods, the Orthogonal Signal Correction (OSC) filters, to be used to remove structured Y-orthogonal variation from X. However, not all structured Y-orthogonal variation needs to be removed, only the irrelevant variation that creates problems for the PLS model (or other regression type methods) should be removed. Otherwise, OSC methods can increase complexity rather than decrease it.

I will demonstrate how a new generic method, Orthogonal Projections to Latent Structures (OPLS) can act both as an OSC-filter and as a stand-alone modelling method with clear similarities with PLS but has one important difference: OPLS (with its built-in OSC filter) detects and removes the structured Y-orthogonal variation in X only when it disturbs the interpretation of the two-block (X-Y) model. OPLS predictions are the same as for PLS, but OPLS separates the relevant and the non-relevant variation in X which improves model interpretation and this makes it so unique