Now that I actually tried to use it, I find the Harmonic class a bit
odd. For one thing the parameter is called "standard deviation" or
something, which doesn't mean anything for a harmonic.
Read the '93 Modeller paper, and perhaps also the appendices to the
Modeller manual, and all will become clear.
Secondly, once
you figure out what it is, it is multiplied by a many significant
figure constant before it is used (the documentation says this is for
modeller compatibility).
It has nothing to do with Modeller compatibility - that's sqrt(RT/2)
where energy units are kcal/mol and T is 297.15K. If anything, that's a
CHARMM/GROMOS compatibility thing. The documentation in the code is
obviously nonsense, and my suspicion is that Bret "derived" this magic
factor empirically, since it differs from the exact value.
Do we really want this? In my mind the thing that makes the most sense
is to to have a the mean and the spring constant be the parameters and
to get rid of the odd multiplier.
The two are equivalent, of course. One is easier to use if you think in
scoring-function space, and the other is easier to use if you think in
coordinate space (e.g. harmonic restraint on a distance with a stdev of
0.1 will result in fluctuations of 0.1 at room temperature). I have no
strong preference for either, since it is very very easy to interconvert
them. Since we have a lot of code that relies on the existing behavior,
we'll stick with it for now unless others would also prefer the kxx/2
form, in which case I'll update the code accordingly.