Despite the advancement in sequencing technologies, around 98% of the genome is usually disregarded due to the lack of interpretation methods. Here, I compare different sequence-based deep-learning approaches for predicting the functionality of the non-coding genome. Using the largest non-coding variant database, I tested the change in prediction as pathogenic vs. benign variants were introduced. Then, I benchmarked their performance on different genomic regions and phenotypes and built a logistic regression model for cell- and phenotype-specific track selection. The models outperformed state-of-the-art evolutionary- and variantbased methods. Finally, I compared different target-gene annotation databases using ontology-based Resnik’s semantic similarity. I combined the previous steps in a variant-to-phenotype or phenotype-to-variant workflow and applied it to rare variants.