Projection to latent correlative structures, a dimension reduction strategy for spectral-based classification

Area of Science:

• Chemometrics
• Spectroscopic Data Analysis
• Bioinformatics

Background:

• Latent variables are essential for dimensionality reduction in chemometrics, particularly with high-dimensional spectroscopic data.
• Common methods like Principal Component Analysis (PCA) and Partial Least Squares (PLS) yield latent variables lacking clear physicochemical interpretation.
• This limitation hinders direct application in understanding underlying chemical or biological processes.

Purpose of the Study:

• Introduce a novel data reduction strategy, Projection to Latent Correlative Structures (PLCS), for chemometrics.
• Develop latent variables with enhanced physicochemical interpretability.
• Validate the PLCS approach using real-world spectroscopic data.

Main Methods:

• PLCS requires a set of reference spectra for defining latent variables.
• Each latent variable quantifies the relative similarity of a sample spectrum to pairs of reference spectra.
• The latent structure is derived from all possible reference spectrum pairings.
• PLCS was combined with soft discriminant analysis for outlier detection.

Main Results:

• The PLCS method was validated on over 500 FTIR-ATR spectra of cool-season culinary grain legumes.
• The approach successfully generated interpretable latent variables.
• The combination of PLCS and soft discriminant analysis proved effective for outlier identification.

Conclusions:

• PLCS offers a valuable alternative for data reduction in chemometrics, providing interpretable latent variables.
• The method shows promise for analyzing complex spectroscopic datasets, especially in agricultural and biological sciences.
• PLCS facilitates deeper analysis by enabling robust outlier detection.