I have been looking into 3D collision detection, but found that it mostly included using bounding boxes, which are not ideal for oddly shape molecules.

Yeah. Sorry about that. The only example I found, which you probably already saw, was more collision avoidance in footballing robots. And all the information I read on that was purely theoretical--no code available that I found. Indeed, pretty much everything I read about SVMs is that way; couched in terms and notations designed to impress peer review boards with little or no information or details on practical implementations.

In general with computation geometry algorithms, if it's useful and can be made fast, then the game writers are your best reference for practical implementations. And understandable descriptions. Either they haven't caught on to SVMs yet, or haven't found a way to make them work quickly enough for their needs.

However, don't be too ready to dismiss bounding. Calculating bounding spheres is well studied, and there are algorithms that give exact accuracy; or very fast approximations; or that can approximate fast and then refine incrementally to meet requirements.

And the nice thing about bounding spheres is that they allow your determination to be made very quickly in many cases--bounding sphere don't intersect--and then allow you to focus on a tiny subset of the points within each groups--perhaps just 1 each--to make the determination when they do intersect.

Regarding my request for a sample dataset--or preferably two sets; one with a clear separation plane and one without: it's just that it is an interesting problem (to me). And whilst I can generate test data for my explorations, it is always better to have real data. Invariably whatever you generate randomly omits some constraint or other that exists in the real problem space.

If the datasets are too large to post I'd gladly fetch them or receive them some other way. I don't need any identifying data, just the points themselves?