On 3/11/11 3:07 AM, David RodrÃguez wrote: > Taking Nick's advice into account, I switched on the dynamic > Lennard-Jones restraints with the simple script shown below. Some models > are generated, showing improved statistics regarding clashes and > stereochemical quality, but many others fail to be generated due to NaN > values in the optimization process
That doesn't sound surprising - LJ will blow up if any pair of atoms get too close together, and many of the randomized initial models are likely to have such clashes. You could building models in the regular fashion and only then refining them with LJ, or modifying the randomization method to avoid introducing clashes. But what you're doing already sounds fine - just build models and discard the ones that fail with NaN.
Ben Webb, Modeller Caretaker