Daniel Russel wrote: > My thought for an improved implementation was to throw the points into a > voxel grid and then have an iterator which traverses the voxel grid and > generates pairs on the fly (rather than generating and storing them).

Well, that sounds essentially like a cell-based nonbonded list, which is the most common good way of implementing a nonbonded list (and is what Modeller does) except for the bit about generating pairs on the fly.

> How far the iterator searches for the second point for each voxel would > be determined by the distance cutoff. It seems easy enough to implement.

The trick is choosing the voxel size sensibly, akin to choosing a hash bin size. This is a bit trickier if you have multiple iterators with different cutoffs.

BTW, another thing to be aware of is that most MM force fields (including CHARMM) exclude not just 1-2 pairs (bonds) but also 1-3 and 1-4. But I guess it would be pretty straightforward to exclude those pairs too by following the bond graph.

Ben