Toward Advanced Quantitative Prediction of Missense Variant Effects

Predicting how missense variants alter protein function is central to resolving the genotype-to-phenotype equation, a task now addressed by numerous computational predictors. The most accurate are increasingly opaque, however, limiting both their mechanistic value and their adoption in the clinic. Here we present QAFIsplit, a redesign of the Quantitative Assessment of Functional Impact framework that models variant effect in two interpretable stages: the mutation tolerance of a position, and the deviation introduced by a specific substitution. Framed as the prediction of a molecular endophenotype, QAFIsplit reaches state-of-the-art quantitative accuracy on independent deep mutational scanning benchmarks, surpassing black-box predictors such as AlphaMissense in most protein domains. Carried into clinical variant classification, it performs on par with established clinical tools without ever being trained on clinical labels, and complements meta-predictors without penalty. In the CAGI 7 community experiment, QAFI-based strategies were submitted blindly to four independent disease-gene challenges, backed by strong pre-submission clinical validation (AUC up to 0.99).