Understanding the functional sequelae of amino-acid replacements is of fundamental importance in medical genetics. Perhaps, the most intuitive way to assess the potential pathogenicity of a given human missense variant is by measuring the degree of evolutionary conservation of the substituted amino-acid residue, a feature that generally serves as a good proxy metric for the functional/structural importance of that residue. However, the presence of putatively compensated variants as the wild-type alleles in orthologous proteins of other mammalian species not only challenges this classical view of amino-acid essentiality but also precludes the accurate evaluation of the functional impact of this type of missense variant using currently available bioinformatic prediction tools. Compensated variants constitute at least 4% of all known missense variants causing human-inherited disease and hence represent an important potential source of error in that they are likely to be disproportionately misclassified as benign variants. The consequent under-reporting of compensated variants is exacerbated in the context of next-generation sequencing where their inappropriate exclusion constitutes an unfortunate natural consequence of the filtering and prioritization of the very large number of variants generated. Here we demonstrate the reduced performance of currently available pathogenicity prediction tools when applied to compensated variants and propose an alternative machine-learning approach to assess likely pathogenicity for this particular type of variant.