Usually 4 is used as threshold. However, it depends from data available and in your possess, from molecules already discovered and your approach to the problem. Moreover if you are doing FBDD you can consider 3. If you use 2 you will have a number of false positive (or negative) due to micellization, precipitation, aggregation and so on.

Hi,

yes, you can be considered to be an inactive ligand 

If you are using this for modeling a QSAR any data that is reliable should be used. Having said that,  I assume you are going to use it a pharmacophore modeling in which you needs a threshold.  A pKi of 2 is definitely inactive in that case, and as mentioned earlier 4 is more normal.   

Hi everyone, from the point of 'does the drug bind' then a pKi of 2 (Ki = 10 mM) can be a very real and specific binding event. There are many published fragments with such weak Kd's that have been converted into much more potent molecules (just take the early stromelysin work for one example, I think that project had a Ki of 17mM). If the Ki is 100mM (pKi 1) then you are really in an extremely weak space where I would say even the intrepid fragment person would not dare tread! Detection methods and limits of solubility are rather challenged down at that level so it may be difficult to prove the binding is real. For your specific studies "discovering features in proteins that are involved in binding active compounds" you should take these 10mM binders seriously. Its best to avoid the term 'inactive', if you have to use it say 'not active at x mM concn'. The question as to what makes a cpd bind at all could reveal better ways to find leads for fragment programs via 'virtual fragment screening'....potentially exciting! Good luck.

Measuring pKi beyond 4 in a biochemical assay reliably (without false positives, see Massimilanos comment) is not straightforward. In our hands NMR more easily allows to go down to 3 (2). I would consider a low pKi (