Relevant molecular descriptors selection

Question

[Vinay]

I have calculated around 1000 molecular structure descriptors (2d and 3d) and fingerprints (such as MACCS, Pubchem fingerprint and Substructure fingerprint) for protein binders from non-binders. I need to create a classifier that could discriminate binders from non-binders. Before that, I need to remove redundant features (descriptors and fingerprints) and reduce data to only significant features those could differentiate binders from non-binders. Now, I have few queries to be addressed:

1. I know non-redundant features could be removed by computing PCC and removing highly correlated features. While computing PCC, is it necessary to treat descriptors and fingerprints independently, or can PCC be computed over the whole dataset?
2. Can Logistic regression modeling (using elastic net) be used for descriptors and fingerprints selection?
3. Other possible methods used for descriptors and fingerprints selection?

All suggestions and helpful comments are appreciated.

[cdsouthan]

Goodness, we don't see many cheminformaticians around these parts so we'll see who chirps up. I wont rush in I'm afraid but the proteochemometrics literature in this area is substantial. But JFTR how are you finding the non-binders ?

[Vinay]

@cdsouthan: I am considering non-binders as decoys of binders, where binders are obtained from literature studies.

[cdsouthan]

Not sure that's a good idea, better to have experimental results rather that to predict non-binding. Why not high vs low affinities? e.g under 100 nM vs over a uM

[Vinay]

@cdsouthan: Since I have IC50 values of compounds available with me, considering high vs low affinities is really a good idea. Thanks for the suggestion.