Study bias inflates cross-phenotype gene overlaps: a macrocephaly case study

Gene sets drawn from phenotype ontologies are widely used to ask how much two conditions share their genetic basis. These comparisons rest on an assumption that is rarely tested: that a gene is equally likely to be annotated to any phenotype. In practice a small number of genes have been studied so intensively that they are annotated to hundreds of phenotypes, while most genes have only a few annotations. Shared membership of these heavily studied genes can create the appearance of a shared genetic basis where none exists. Using macrocephaly as a case study, I built gene sets for macrocephaly, microcephaly, autism and seizure from the Human Phenotype Ontology and measured their pairwise overlaps with the standard hypergeometric test. All three overlaps with macrocephaly were strongly significant (1.5–2.2 fold, p < 1e-4). I then retested each overlap against a permutation null matched to each gene’s annotation load, so that the comparison asks whether the overlap exceeds what an equally well-studied random gene set would give. Under this test the macrocephaly– microcephaly overlap fell to 1.00 fold (p = 0.51) and the macrocephaly–autism overlap to 1.07 fold (p = 0.29); only the macrocephaly–seizure overlap remained significant, and modestly so (1.14 fold, p < 1e-4). The 37 genes shared by all four phenotypes, which include several of the most heavily annotated genes in the ontology, were also consistent with study bias alone (1.03 fold, p = 0.45). These results show that ontology-based overlap statistics can be dominated by how much each gene has been studied, and that a study-bias-matched null is needed before an overlap is interpreted as biological.