Moose is a novel neuro-symbolic learning method for latent concept learning in $mathcal{EL}^{++}$ ontologies. Existing neuro-symbolic learning methods mainly focus on propositional theories or Datalog, while reasoning-shortcut (RS) awareness has not been explored in ontology settings. Moose compiles an $mathcal{EL}^{++}$ TBox and finite ABox to a Sentential Decision Diagram (SDD), which acts as a differentiable weighted-model-counting layer. To overcome the limited expressivity of $mathcal{EL}^{++}$ under partial supervision, Moose adds closure clauses outside the $mathcal{EL}^{++}$ profile on declared exhaustive families. The results show that Moose outperforms traditional propositional neuro-symbolic learning, fuzzy logic, and ontology embedding baselines on MNIST-with-ontology and Pizza"iolo datasets, and presents the first reasoning-shortcut analysis in an OWL EL setting. Blogger's Review: Moose provides a new perspective for latent concept learning in ontologies, its reasoning-shortcut awareness and differentiable weighted-model-counting layer enable it to perform better and be more flexible in complex ontology environments.