I've been thinking of hooking up the new Accelrys QSAR Workbench to ChEMBL for a while now, and I've made some progress in this direction. Please see the attached Word document, which describes my initial efforts.
As a quick summary, I've found that for certain assays, with no missing or erroneous values, the QSAR Workbench provides a good way to quickly build models. It is possible to quickly (and combinatorially) explore different training/test set splits, descriptor subsets, and learners and learner parameters. I've added a comment to the 'Predictions' discussion, with an attachment with some Bayes models for one such assay.
For other assays, however, there is some data cleansing which will need to happen, and this hasn't been implemented in the QSAR Workbench yet (though initial steps towards this have been made by my colleague Willem van Hoorn). As this simply consists in writing protocols, it shouldn't be hard to do. I'll update this discussion once I've made some progress.
Finally, it should be simple to cannibalize the integration to pull out a few "generic" PP components (e.g. ChEMBL Assay to SD File), and I'm going to be working with Mike Cherry to do this in the near future.