Energy minimization of my system with both steepest descent and conjugate gradient converges very quickly – under 100 steps – with the following (typical) result:
Steepest Descents converged to machine precision in 62 steps,
but did not reach the requested Fmax < 100.
Potential Energy = -4.8300015e+06 Maximum force = 5.6579081e+05 on atom 55124
Norm of force = 1.4625802e+03
The exact values vary run to run, but max force is always over 1e5. The atom isn’t always the same, but it’s always in the same side chain. So it would appear that side chain is stuck in a very energetically unfavorable local minimum, and downhill minimization is unable to find a way out of it.
Can anyone recommend a protocol that could “jiggle” the systems out of this energy minimum? Something that allows uphill steps, or maybe slowly ramping up the repulsive LJ component? Or is rebuilding the system from scratch in hopes of avoiding this clash likely to be the simplest option?
Hi, this isn’t going to be particularly satisfying but when I encounter this problem I minimizing the structure in namd. The minimization protocol here is pretty robust and doesn’t usually give me greif because it uses conjugate gradient decent. There is this option in gromacs too but it complains if you don’t do steepest decent first.
Beyond this you can try to resolve the clash yourself by tweaking coordinates by hand (tedious).
Before trying anything I would have a good look at the system to make sure everythign is ok. Sometimes for example charmm-gui will just put two hydrogens in a lipid bilayer too close together.
Yes, I was hoping a really clever protocol is out there that might allow me to avoid having to fix this “by hand”, but it was probably too much to hope for. Examining the system visually makes the problem quite clear: the tool I used to model in missing residues made a particularly bad choice in the vicinity of the residue that ends up being “stuck”. And even though the solution is immediately clear “by eye” to the human brain (building in several terminal residues extended the backbone through a neighboring loop, essentially knotting the backbone), it requires the kind of non-local adjustment (“unknotting”) that seems almost unfair to ask of an algorithm.
Sigh … yet again this makes me wish for a FoldIt-like interface that would allow manipulating the polypeptide chain with the mouse.