Multivariate unmixing approaches on Raman images of plant cell walls: new insights or overinterpretation of results?

Background: Plant cell walls are nanocomposites based on cellulose microfibrils embedded in a matrix of polysaccharides and aromatic polymers. They are optimized for different functions (e.g. mechanical stability) by changing cell form, cell wall thickness and composition. To reveal the composition of plant tissues in a non-destructive way on the microscale, Raman imaging has become an important tool. Thousands of Raman spectra are acquired, each one being a spatially resolved molecular fingerprint of the plant cell wall. Nevertheless, due to the multicomponent nature of plant cell walls, many bands are overlapping and classical band integration approaches often not suitable for imaging. Multivariate data analysing approaches have a high potential as the whole wavenumber region of all thousands of spectra is analysed at once.

Results: Three multivariate unmixing algorithms, vertex component analysis, non-negative matrix factorization and multivariate curve resolution–alternating least squares were applied to find the purest components within datasets acquired from micro-sections of spruce wood and Arabidopsis. With all three approaches different cell wall layers (including tiny S1 and S3 with 0.09–0.14 µm thickness) and cell contents were distinguished and endmember spectra with a good signal to noise ratio extracted. Baseline correction influences the results obtained in all methods as well as the way in which algorithm extracts components, i.e. prioritizing the extraction of positive endmembers by sequential orthogonal projections in VCA or performing a simultaneous extraction of non-negative components aiming at explaining the maximum variance in NMF and MCR-ALS. Other constraints applied (e.g. closure in VCA) or a previous principal component analysis filtering step in MCR-ALS also contribute to the differences obtained.

Conclusions: VCA is recommended as a good preliminary approach, since it is fast, does not require setting many input parameters and the endmember spectra result in good approximations of the raw data. Yet the endmember spectra are more correlated and mixed than those retrieved by NMF and MCR-ALS methods. The latter two give the best model statistics (with lower lack of fit in the models), but care has to be taken about overestimating the rank as it can lead to artificial shapes due to peak splitting or inverted bands.