I have a data set of peak areas from gas chromatography I would like to run on a PLSR model. Generally, for PLSR I would center and scale the data, is that appropriate here?
As the peaks differ in scale between compounds on a magnitude of 100, running the model on unscaled data is unfeasible.
Is it standard to scale these peak areas? Is there a scaling method that will reduce overfitting the model and avoid introducing extra noise?
Firstly, you must perform scaling before you perform a PLSR or PLS-DA, PCA model. Some compounds are dominant in concentration in your sample so their peaks usually have higher peak intensity and variances. Therefore, if you don't scale your data, peaks with high intensity will be recognized as the most important differential metabolites and you will lose the information of others.
For scaling, you can choose:
1. Autoscaling
2. Range scaling
3. Parero scaling
These three scaling methods were often chosen. For more details about the advantages and disadvantages you can check the paper written by Robert A van den Berg at Article Van den Berg RA, Hoefsloot HCJ, Westerhuis JA, Smilde AK, Va...
although the paper was mentioning the peak table from GC-MS or LC-MS, it is the same issue as yours, the features in MS datable are m/z, in your case is rt.
Secondly, as for reducing overfitting, you can compare the consequences of accuracy when you choose different scaling methods, but more importantly, is usually the overfitting is related to the shape of your dataset. The best way to solve overfitting is to increase the number of samples. In PLSR, there is one way to reduce overfitting is to reduce the number of components you use to do prediction.
To avoid extra noise, once you do scaling, no doubt that all the noise will be amplified, remember all the scaling methods aim to regard all metabolites as having the same importance. But Parero scaling may work a little better than others. To reduce the noise, the best way is to remove the noise before scaling. Usually, we utilize QC samples. All the peaks in your QC samples with variances higher than 20% (or another number, please check papers recently published, 20% is usually used in MS data) will be recognized as noise and unstable metabolites and removed from the dataset.
Hope this will be useful for you.
- Badges
- Science topic
- Similar topics
- Methodology
- Research Methodology
- Research Papers