Background: Predicting the deleteriousness of observed genomic variants has taken a step forward with the introduction of the Combined Annotation Dependent Depletion (CADD) approach, which trains a classifier on the wealth of available human genomic information. This raises the question whether it can be done with less data for non-human species. Here, we investigate the prerequisites to construct a CADD-based model for a non-human species. Results: Performance of the mouse model is competitive with that of the human CADD model and better than established methods like PhastCons conservation scores and SIFT. Like in the human case, performance varies for different genomic regions and is best for coding regions. We also show the benefits of generating a species-specific model over lifting variants to a different species or applying a generic model. With fewer genomic annotations, performance on the test set as well as on the three validation sets is still good. Conclusions: It is feasible to construct species-specific CADD models even when annotations such as epigenetic markers are not available. The minimal requirement for these models is the availability of a set of genomes of closely related species that can be used to infer an ancestor genome and substitution rates for the data generation.