Genotype-phenotype relationships in phenylalanine hydroxylase deficiency: Functional annotation-enhanced analysis of 23,427 individuals.

Abstract

Phenylalanine hydroxylase deficiency spans from mild hyperphenylalaninemia (MHP) to mild PKU phenylketonuria (mPKU) and classic PKU (cPKU). Genotype-phenotype inference is complicated by allelic heterogeneity and the incomplete functional annotation of complementary DNA-only variant strings. We analyzed 23,427 individuals with 2 PAH alleles and metabolic phenotype (MHP [n = 4208, 18.0%], mPKU [n = 5295, 22.6%], and cPKU [n = 13,924, 59.4%]; 10,108 [43.2%] individuals had blood phenylalanine values). Variants were functionally annotated with Ensembl variant effect predictor and SpliceAI and mapped to 3 functional classes: predicted loss-of-function, splice-uncertain, and missense/other. We quantified genotype-phenotype concordance and evaluated phenotype prediction using ordinal and multinomial models. Variant effect predictor provided functional consequences for 1007 unique variants and annotated >99% of alleles. The genotype functional class showed a strong relationship with phenotype, with 0/0 genotypes predominantly classified as cPKU. Genotype-phenotype concordance increased with genotype frequency, and common genotypes displayed high phenotype consistency. An ordinal ridge model using allele identity plus functional class achieved an accuracy of 0.790 (quadratic weighted κ = 0.784) under genotype-held-out evaluation. A multinomial logistic model achieved an accuracy of 0.836 on a random patient split. Continuous Phe prediction using ridge regression on log(Phenylalanine) achieved an R² of 0.673 with a mean absolute error of 357 μmol/L. Benchmarking against the published allelic phenotype value/genotypic phenotype value system yielded an accuracy of 0.849 in 22,656 individuals with allelic phenotype values for both alleles; performance was high for cPKU and MHP but lower for mPKU, consistent with prior reports. In this large cohort, PAH genotype is strongly associated with metabolic phenotype. Functional consequence annotation enables mechanistic interpretation (loss-of-function and splice effects) and improves the portability of genotype-based predictions to previously unseen genotypes.