Groups are a dead end, since there is no way to restrict or differentiate pairwise interactions by groups. Thus I think the only way that you can currently realize such a model would be to have different atom types for each molecule and then use the hybrid pair style. So if your molecule has, say, two atom types, you would have to have 200 atom types for a 100 molecule system and then you can a define different pair style for interactions between types 1 and 1, 1 and 2, or 2 and 2 compared to 1 and 3, 1 and 4 etc. This may be a bit tedious, but it can be easily scripted and in the pair_coeff command you can use wildcards for ranges of atom types.
Read more here: Source link