Skip to content

S-Plot Marker Plot

🎯 What is this?

S-Plot is the "killer" chart of the OPLS family (OPLS / OPLS-DA), specially used to screen markers (Biomarker / Marker).

It answers two questions at the same time:

  1. Is this variable's contribution large?
  2. Is this variable trustworthy?

An ordinary correlation analysis can only answer one of them. S-Plot draws both on the same chart, so variables that are "both important and reliable" automatically gather in two corners.


🧐 How to read?

AxisMeaning
X-axis (p1)Covariance——the variable's contribution to the model
Y-axis (p(corr))Correlation coefficient——the reliability of the correlation between the variable and the result

Meaning of the four regions:

PositionMeaning
Top right cornerLarge contribution and stable positive correlation → positive marker candidate
Bottom left cornerLarge contribution and stable negative correlation → negative marker candidate
Middle rightLarge contribution but unstable correlation → suspicious result, use with caution
Near the originSmall contribution → a variable that can be removed

💡 One-line memory aid: the closer to the corner, the more reliable. True markers "swing" out to the top-right or bottom-left corners, forming an obvious S shape (which is also where the name S-Plot comes from).


🛠️ How to use?

1. Quickly lock onto key variables

Use the lasso to circle the top-right / bottom-left region; these variables are the marker candidates to focus on first.

2. Simplify the model

Remove the variables near the origin and rebuild the model; this usually makes the model more concise with better generalization.

3. Cross-validate with VIP

ChartFocus
VIP PlotOnly "contribution size" (one-dimensional ranking)
S-PlotBoth "contribution + reliability" (two-dimensional)

💡 Best practice: take the intersection of the two——variables that are both on the VIP > 1 list and fall in the corners of the S-Plot are the truly reliable markers.


⚠️ Applicable models

S-Plot is available only for the following models:

  • OPLS / OPLS-DA (recommended; after orthogonal correction the covariance is cleaner)
  • PLS / PLS-DA (usable, but lacking orthogonal noise stripping, the pattern may be more divergent)

⚠️ PCA models have no Y, so covariance and correlation cannot be computed, and S-Plot is not supported.


🎯 Typical applications

  • Authenticity identification: find the characteristic wavelengths that distinguish genuine products from imitations in spectral data
  • Raw material grading: locate the key physicochemical indicators that determine the grade
  • Metabolomics: screen differential metabolites
  • Process diagnosis: find the key parameters that distinguish "normal batches" from "abnormal batches"

💡 After finding candidate markers, it is recommended to use Causal Inference (DML) to further verify the causal relationship, and avoid mistaking "accompanying changes" for "adjustable levers".

Let data speak, make decisions simpler.