Finding most discrminatory / influential proteins to discriminate datasets

Question

[The]

Hello All,

I’ve two proteomics datasets, one consists of biological replicates for responders ( to a particular drug) and the second one consists of biological replicates of non-responders to the same drug. For quantitative values, I’ve intensities measured for all proteins common to these datasets.

I wish to know which proteins are the most influential ones in discriminating these two datasets. I’ve tried PCA, however the results are not very good. I’m particularly looking for ‘better’ methods particularly machine learning based ones. Can anybody suggest any such methods?

Thanks in advance

Answer 1

[Jean-Karim Heriche]

You could try linear discriminant analysis.

Answer 2

[Santosh Anand]

Random Forest and Support Vector Machine are two really good binary classifiers if you have large number of samples (a few hundreds?). You must do it properly and honestly to get the real clinically relevant results. Like any other non-linear method, RM and and SVM can easily overfit any kind of data. You should have proper training, discovery (and if possible validation) sets of data to have good faith in results. One of the papers that I remember did it very well (and explained well too) is

Mass Spectrometry to Classify Non–Small-Cell Lung Cancer Patients for Clinical Outcome After Treatment With Epidermal Growth Factor Receptor Tyrosine Kinase Inhibitors: A Multicohort Cross-Institutional Study

Although it uses a different classifier (KNN), the main idea remains the same. There are many other paper you can find on pubmed (these algorithms are quite common in biomarker field).

If you are looking to get your hands dirty right away, try RapidMiner or Weka. They have very easy to use implementation of these algorithms and many more.