Variable locus length in the human genome leads to ascertainment bias in functional inference for non-coding elements

Motivation: Several functional gene annotation databases have been developed in the recent years, and are widely used to infer the biological function of gene sets, by scrutinizing the attributes that appear over- and underrepresented. However, this strategy is not directly applicable to the study of non-coding DNA, as the non-coding sequence span varies greatly among different gene loci in the human genome and longer loci have a higher likelihood of being selected purely by chance. Therefore, conclusions involving the function of non-coding elements that are drawn based on the annotation of neighboring genes are often biased. We assessed the systematic bias in several particular Gene Ontology (GO) categories using the standard hypergeometric test, by randomly sampling non-coding elements from the human genome and inferring their function based on the functional annotation of the closest genes. While no category is expected to occur significantly over- or underrepresented for a random selection of elements, categories such as 'cell adhesion', 'nervous system development' and 'transcription factor activities' appeared to be systematically overrepresented, while others such as 'olfactory receptor activity'-underrepresented. Results: Our results suggest that functional inference for non-coding elements using gene annotation databases requires a special correction. We introduce a set of correction coefficients for the probabilities of the GO categories that accounts for the variability in the length of the non-coding DNA across different loci and effectively eliminates the ascertainment bias from the functional characterization of non-coding elements. Our approach can be easily generalized to any other gene annotation database.