How GeneMANIA works is described in the FAQ and this paper or look here for a slightly different perspective. Basically, what it does is derive a measure of functional similarity between genes from the network structure. Given a query, the other genes are simply ranked by their similarity to the query. This works better for small gene lists as this can exploit the network structure. With a single query gene, you're not really taking advantage of the network as this basically looks at the nearest neighbours.
By selecting different sources, you're building a different network and depending on this network, your query may not be connected to some other genes hence the different results. On top of this, GeneMANIA uses weights to maximize connectivity between the query genes or, for small gene lists, to maximize connectivity of genes annotated with the same GO biological process terms as the query genes.