Resolving Protein Identification Conflicts

Question

[Graslevy]

Hey Guys, I am trying to solve the problem of resolving protein identification conflicts. After searching Mascot to identify peptides, some peptides are often identified as evidence for various/multiple proteins (some due to homology, others due to errors in DB); the task is to resolve such conflicts by excluding false positive identifications.

In my case, I have proteomics data from C. difficile. To carry out a logistic regression analysis, I have ascribed peptides mapped to proteins from this organism as TRUE and others as FALSE. I am achieving ~94% prediction accuracy overall, with half the false positives correctly excluded..this is unsatisfactory.

Please highlight any problems with my approach and can you suggest alternative approaches to solving this problem? I am aware of the PeptideProphet solution (described mainly for SEQUEST data) but will like a different approach.

Answer 1

[Graslevy]

I acknowledge there is no easy solution to this problem and some conflicts will persist. I had to choose between using a supervised learning solution (the one described in the question) and a soft classifier (as previously described by others including PeptideProphet). I ended up implementing an expectation maximization classifier; although this solution excludes more false positives, it inadvertently excludes ~20% of the true positives. Compared to the other approach, prediction accuracy is less but it is more dynamic and adaptive.

In the absence of any suggestions, I settled for the soft classifier.