Cosmological simulations aim to understand the matter distribution in the universe either by following a semi-analytic approach or by formulating a hydrodynamical version of the matter distribution. Both approaches describe the evolution of baryonic structures inside potential wells created by dark matter, while the dark matter itself is modelled as self gravitating collision-less system. While major advancements have been made to reduce the computational costs, these simulations still take millions of CPU hours to converge to a stable set of solutions. This naturally leads to the question: can generative models predict the properties of a galaxy, given a partial history of its dynamical evolution? Given that computing conditional probabilities is intractable in general, tractable probabilistic models such as sumproduct networks have emerged, where conditional marginals can be computed in time linear in the size of the model. In this work, we investigate how sum-product networks can be used to compactly represent and learn distributions for prediction in concordance cosmology. Our results study the extent to which they can infer the relation between baryonic matter and dark matter. We test the algorithm on the Eagle suite of cosmological hydrodynamical simulations to show that graphical models can satisfactorily reproduce mock catalogs of galaxies.