stevesmedchem Registered
I've downloaded and tried out the HLM model that gave the best results when looked at the three protocols used for model forming.
I've tried running this on several data sets and the results look possible for the small number of compounds I checked by eye. However, there are many with the 'inf' marker in the statistics column.
The sets I have are reasonably 'drug-like' with most having a decent mix of MWt/logP/TPSA/structural features, so I would concur with the comment that this model should be valuable for those compounds close to the training set (30-50% of the ~ 35K compounds I tried). Based on my results, many compounds are likely to fall outside the compound space for the model.
The model ran quickly, so it's easy to check the situation with your own set of compounds.