In general, the experiments we do (virtual or real) are designed to answer a research question or to test a research hypothesis. Accordingly, the compounds you include in your QSAR models depend upon the question you are asking or the hypothesis you are testing.

Usually, the place to start when attempting to make predictions of toxicity is with one compound at a time. If the impurities that are listed in the various compendiums appear to be of interest, you could make predictions of each these compounds separately. Then, a more difficult question would be to assess the toxicity of mixtures of the principal compound along with its impurities.

There are some specific cases in which the presence of certain impurities is known to be quite important. For example, in early tests of the in vitro inhibitory potency of the insecticide parathion against acetylcholinesterase, it was found that the compound possessed a rather high potency. Later, it was discovered that the parathion that was used in the initial experiments contained an impurity, paraoxon (the active metabolite of parathion), which has a very high potency against acetylcholinesterase. Now it is accepted that parathion itself has a comparatively low in vitro potency for inhibiting acetylcholinesterase.

So, in your toxicity predictions, it would be a matter of your professional judgment whether or not to assess the toxicity of the impurities that are known to coexist with your main compounds of interest.

Nah, usually not—if the impurities are already listed in USP or EP monographs, they're considered safe within those limits, so QSAR isn't required unless levels exceed specs, it's a new context, or regulators flag genotoxicity concerns per ICH guidelines. Check ICH M7 for edge cases.