Assessing Computational Models for Pharmacogenomic Variant Interpretation

Effectively anticipating the impact of pharmacogenomic variants is crucial for creating tailored therapeutic approaches, given the fact that genetic variation can lead to differing drug responses among individuals. In this study, researchers evaluated multiple computational techniques using a collection of pharmacogenomic variants with clinical annotations or functional insight from deep mutational scanning, sourced from existing literature. The team also concentrated on CYP2C9, a drug-metabolizing enzyme of notable clinical importance.

The findings revealed that, despite recent technical advancements, there is still significant room for enhancement. Specifically, existing models encounter difficulties in differentiating gain-of-function variants linked to amplified drug clearance and fast-metabolizer traits from neutral variants. Conversely, loss-of-function variants that diminish drug clearance are predicted with greater accuracy.

The incorporation of structural and evolutionary data appears to be crucial for enhancing performance, with the coevolution-based StructureDCA method attaining the highest accuracy among classical genetic variant-effect predictors and recent deep learning methods, including the AlphaMissense pathogenic-variant predictor and protein language model-based systems.

Moreover, the study indicates that computational models can serve as valuable supplementary tools to in vitro experiments in clinical variant interpretation. The StructureDCA predictions demonstrated a stronger alignment with clinically annotated phenotypes compared to extensive deep mutational scanning data in several instances.