PRM Proteomic Data

Okay, I have spent a whole day now looking for an answer. I think I have a unique case when it comes to scaling and SVM. Briefly, the data attached is proteomic data collected from mass spectrometry. Each row is a protein and columns (mEV_1....nEV1...etc) are different donors/blood samples. Each protein has been normalized to internal control (a specific protein) in order to calculate a relative amount (number in each cell).

There are obvious differences in the range of numbers within each protein, however, they aren't ridiculously different like example online. My main question is if it is always necessary to scale data for rbf SVM? I understand one feature may have more influence, however, this may be the case because not every protein will be at equal levels in the blood. The rbf SVM I have generated is able to get 100% accuracy on a 70/30 train/test split with or without scaling (albeit different hyperparameters were used).

  • $\begingroup$ You are getting 100% accuracy both in the training and test set? $\endgroup$ – Johanna Aug 28 '20 at 14:25
  • $\begingroup$ 100% accuracy with or without scaling the data. $\endgroup$ – thejahcoop Aug 28 '20 at 14:28
  • $\begingroup$ Yes, but are you getting 100% accuracy on an independent test set? Because if not, it is very likely that your model overfit your data $\endgroup$ – Johanna Aug 28 '20 at 14:32
  • $\begingroup$ Yes, i have 70%/30% train test split coded in. I am mostly asking at what point do we consider a range of values to be "different" from each other.... or is it better to just scale the data regardless of the application. $\endgroup$ – thejahcoop Aug 28 '20 at 14:36
  • $\begingroup$ I think I answered my own question, I am going to repeat with scaling as I am sure this will make the peer-review process less inflammatory. $\endgroup$ – thejahcoop Aug 28 '20 at 14:44

It essential to scale data, yes. Even if your data doesn't have a thousand integers different, it would be advisable to normalize it in the [0,1] interval, for example.

You could be getting the 100% accuracy with or without scaling due to, for example, having a very skewed training and test set; your test data could have observations too similar to your training data (some kind of repeated observations)...

How did you perform your scaling? Did you normalize each feature with the max/min method?

  • 1
    $\begingroup$ Thanks! I had to take a closer look at the rbf funcion to see why this was the case. I didn't scale to begin with because the values of each protein was normalized to a protein standard injected with each sample. This control large extremes, but as you mentioned, still produces a different range of values. Unfortuantely, I have to remake the figures, but less of a headache than a bad peer-review. Cheers. $\endgroup$ – thejahcoop Aug 28 '20 at 15:09

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.