Dimensionality reduction & clustering / (O)PLS-DA
Your data stays in your browser: files are read and processed on your device, never uploaded to any server, and cleared when you close the page. Free, no sign-in.
PLS-DA is supervised dimension reduction: it looks for the directions that most clearly separate the known groups, so separation in the score plot is expected by design and does not by itself show that the groups differ.
Look at the cross-validated Q² (closer to 1 is better; above about 0.5 is usually considered good predictive ability, ≤ 0 means no predictive value) and its gap from R²Y: a high R²Y with a low Q² indicates overfitting, particularly when variables far outnumber samples.
OPLS-DA puts group-related variation into a single predictive component (x-axis) and systematic variation unrelated to the groups into orthogonal components (y-axis), which eases interpretation; its predictive ability equals that of PLS-DA with the same number of components.
VIP (variable importance in projection) summarises each variable’s contribution to the model; VIP > 1 is the customary screening threshold. Confirm candidate variables with univariate tests (e.g. t-tests with FDR correction).
Metabolomics data are commonly Pareto- or unit-variance-scaled. Cross-validation for Q² assigns sample i to fold i mod k, so results are reproducible.
Wold S., Sjöström M. & Eriksson L. (2001) Chemom Intell Lab Syst 58:109–130; Trygg J. & Wold S. (2002) J Chemom 16:119–128 (OPLS); Chong & Jun (2005) Chemom Intell Lab Syst 78:103–112 (VIP). PLS-DA scores and weights match scikit-learn PLSRegression on one-hot-coded groups; OPLS-DA matches an independent NumPy implementation of the Trygg–Wold algorithm.
Up to 20,000 samples or variables (large matrices take a few seconds).
Send us your data and research question and you will receive a written plan within 1 working day: analysis steps, parameter rationale, deliverables and timeline. Quoted per project.