It all depends on your research question i.e. what do you want to answer with the DCM. If you are interested if the connections are forward or backward, than you might model 4 possibilities (no modulatory, modulatory on forward, modulatory on backward and modulatory on both), and compare the outcome using model selection.
But I assume that you are not sure if you may question modulatory on backward if it is not on forward? I believe you may.