What I’d like to see are efficient ways to map numeric data back to structures. So rather than just do the usual procedures like MDS etc, can we represent the information in the numeric data in terms of relationships between structure and data (and not just low-D representations of the data).

Ok this is probably too vague! An example is an application at http://cheminfo.informatics.indiana.edu/~rguha/code/java/salivis/ which represents SAR information in the form of networks.

In terms of mol vis itself, I’d like to see useful ways of mapping things to structures themselves - so fragment properties could be mapped to specific parts of a molecule, effects of descriptors in a QSAR model could be mapped to parts of a molecule and so on. Some of this has been done before, but is usually quite restricted.

I suppose one blue-sky thing I’d like to see is a virtual reality style exploration of arbitrary chemical spaces - conceptually, not too difficult, especially since chemical spaces have been discussed before (cf ChemGPS by Oprea), but an efficient and intuitive tool would be great.

Also, you might want to take a look at Ben Shneiderman of U Maryland - he has some extremely cool and state of the art visualization techniques for large, high-D datasets