Sarah Nexus does not take into account 3D conformations. Local structural motifs that can block the activity of reactive fragments through stereo-electronic interactions will be learnt by the model if sufficient data is available in the training set. This will be captured by large fragments that encompass both the activating fragment and the deactivating adjacent groups.
More generally, while it could be possible for a 3D conformation to block a chemically reactive fragment though shielding by non-adjacent groups (e.g. during hydrophobic collapse), we believe that this is unusual. If you do have specific examples we would be really interested to see them; currently they would be great examples for when an expert should use their knowledge to over-rule a prediction.