Abstract
Computational prediction yields efficient and scalable initial assessments of how variants of unknown significance may affect human health. However, when discrepancies between these predictions and direct experimental measurements of functional impact arise, inaccurate computational predictions are frequently assumed as the source. Here, we present a methodological analysis indicating that shortcomings in both computational and biological data can contribute to these disagreements. We demonstrate that incomplete assaying of multifunctional proteins can affect the strength of correlations between prediction and experiments; a variant's full impact on function is better quantified by considering multiple assays that probe an ensemble of protein functions. Additionally, many variants predictions are sensitive to protein alignment construction and can be customized to maximize relevance of predictions to a specific experimental question. We conclude that inconsistencies between computation and experiment can often be attributed to the fact that they do not test identical hypotheses. Aligning the design of the computational input with the design of the experimental output will require cooperation between computational and biological scientists, but will also lead to improved estimations of computational prediction accuracy and a better understanding of the genotype-phenotype relationship.