Applicability domain for multitask learning

Question

Given a multitask machine learning model with several endpoints, I want to determine whether an unseen datapoint is in the applicability domain. I will calculate the tanimoto similarity coefficient and SALI index. My question is, when I calculate these, do I consider my entire training data, or only the part of the dataset that is the endpoint for the main task?

I tried looking at articles, and the QMRF of multi task models, but did not find info about my question.