In short, the effective observation to parameter ratio improves by ~4%. This seems like a relatively small improvement, especially if the trade-off is that Ramachandran statistics can't be used for validation anymore. It also seems like the improvement would decrease with larger proteins (the number of additional parameters from adding a residues increase faster than the number of secondary structure restraints that residue could provide).
There is nothing preventing you from using Ramachandran statistics for validation if the secondary structure restraints only restrain the distances between hydrogen-bonded atoms. (This is what the implementation of S.S. restraints in phenix.refine does.) This actually contributes significantly fewer than two extra restraints per residue on average. And it doesn't actually have much of an effect on Ramachandran statistics; the number of outliers is slightly reduced, as is the clashscore. I suspect (although I haven't examined this in detail) that this is mostly keeping the ends of helices and sheets from unraveling, which is what they're most useful for. (The change in R-factors is negligible in our tests, although they sometimes reduce overfitting a little bit.)
This is an entirely different issue from Ramachandran (i.e. phi/psi) restraints, which should only be used as a last resort at low resolution at the end of refinement, and after fixing all outliers manually in Coot. (Restraining to a high-resolution model, if one is available, is a much better option.)
-Nat
