Dear Abhay Vincent

I would say that it largely depends on what you want to do. For example, do you want to perturb the system or to drive a near-equilibrium transformation?

This being said, looking for previous works done in the same field is quite mandatory to see what has been done, what constants have been used and so on. If you want to unfold a protein, you will need to use force constants large enough to actually unfold it, and not play a tug-of-war with the intra-protein forces of the system. On the other hand, if you want to drag a solute/permeant through a membrane/a protein channel/a nanotube/etc. you may want to not perturb/disrupt the bilayer, or you may want to sample the interaction along the protein channel therefore you need slow velocities to let the permeant interact with its surroundings.

Overall, apart from the limits intrinsic to the software you are using and to the artifact you can introduce, it greatly depends on the theory you are applying. Does it depend on the velocity/k constants? Are you sampling states which need equilibrium conditions, and therefore must perturb the system the least possible? Or does it depend on fast dragging with ensembles of out-of-equilibium trajectories? Does it depend on any approximation, like stiff-spring which maybe would work better with constraints rather than restraints? Or maybe are you trying to reproduce a pulling AFM experiment, and therefore can get some insights from the experimental work?

In a way, it is always a matter of trial and error and checking was has already been done.

Hope this helps,

Nicola