Even carefully designed assays may be blind to some important functional effects. For example, we recently used a generalizable protein abundance assay to identify possible pathogenic variants of PTEN and TPMT . Known pathogenic active site variants, which damage PTEN enzymatic function but do not affect protein abundance, are not identified by this assay. The inability of this assay to identify all pathogenic variants is not an indication that the assay results are unreliable and unsuitable for clinical use. Nevertheless, a clear explanation of this limitation was essential to ensure that downstream users of the data understand what the assay did and did not measure. Even though a generalizable assay like the protein abundance assay cannot measure all functions, it can still be extremely powerful because a significant portion of damaging variants affect gene expression or stability and, therefore, all protein functions [14, 31]. However, an assay that more completely or specifically represents gene function and disease context may be useful in facilitating a more confident prediction of a benign effect, rather than in predicting pathogenic variants more accurately. Thus, for every MAVE, there should be a clear statement of what was measured and a disclosure of the types of functional effects that were not measured.
We recommend that functional assays should only be used once as a pathogenic supporting, moderate, or strong criterion. This is because, in general, it will be nearly impossible to determine whether the separate MAVEs are measuring orthogonal or overlapping functions. Concordance between two functional assays for a variant does add reassurance that functional scores are robust. Machine-learning models can be used to incorporate functional scores from several different assays that were applied to the same functional element to predict an overall variant effect. Initial efforts combining data from multiple MAVEs that query different functions of the same protein have shown that the prediction of pathogenic variants can be improved by integrating the datasets [7, 11, 16]. The improvements in predictive performance are likely to come from the formation of a more complete picture of variant effect. These composite scores should only be used once and should not be stacked with scores derived from individual functional assays, or from orthogonal assays that are not included in the composite score. 2b1af7f3a8