Statistical testing in gene transcriptomic-neuroimaging associations: an evaluation of methods that assess spatial and gene specificity

Multiscale integration of gene transcriptomic and neuroimaging data is becoming a widely used approach for exploring the molecular underpinnings of large-scale brain structure and function. Proper statistical evaluation of computed associations between imaging-based phenotypic and transcriptomic data is key in these explorations, in particular to establish whether observed associations exceed 9chance level9 of random, non-specific effects. Recent approaches have shown the importance of spatial null models to test for spatial specificity of effects to avoid serious inflation of reported statistics. Here, we discuss the need for examination of the second category of specificity in transcriptomic-neuroimaging analyses, namely that of gene specificity, examined using null models built upon effects that occur from sets of random genes. Through simple examples of commonly performed transcriptomic-neuroimaging analyses, we show that providing additional gene specificity on observed transcriptomic-neuroimaging effects is of high importance to avoid non-specific (potentially false-positive) effects. Through simulations we further show that the rate of reported non-specific effects (i.e., effects that are generally observed and cannot be specifically linked to a gene-set of interest) can run as high as 60%, with only less than 5% of transcriptomic-neuroimaging associations observed through ordinary linear regression analyses showing spatial and gene specificity. We explain that using proper null models that test for both spatial specificity and gene specificity is warranted.