The GPCR^VP score reliably predicts the impact of GLP1R human variants on receptor function

The glucagon-like peptide-1 receptor (GLP1R) is a key regulator of glucose homeostasis and body weight, and a major therapeutic target for type 2 diabetes and obesity. Individual disease risk and response to treatments may vary widely depending on the presence of missense variants (MVs) in the GLP1R gene, but accurate prediction of the impact of MVs in the G protein-coupled receptor (GPCR) remains challenging. Here we aimed to generate a reliable and GPCR-tailored predictor of the impact of MVs that is applicable for precision medicine.

We developed a GPCR variant impact predictor (GPCR^VP score) by combining molecular dynamics (MD) simulations with available machine-learning/AI-based models (AlphaMissense and REVEL) to substantially improve the predictive performance of existing models. MD simulations of wild-type and mutant GLP1R complexes were used to extract GPCR-specific structural and dynamic features that capture membrane context, allosteric communication, and mutation-induced contact rearrangements. These descriptors were integrated with existing variant-effect predictors (AlphaMissense and REVEL) and trained on 58 experimentally characterised GLP1R variants to develop GPCR^VP, a GPCR-specific variant predictor.

The GPCR^VP predictor showed strong agreement with experimental data and robust performance on independent validation datasets. Notably, it demonstrated a substantial reduction in false-positive predictions (approximately 78%) in the validation sets compared with AlphaMissense, indicating improved specificity and reliability. Generation of structural models for all known human GLP1R MVs (174 outside the training and validation sets) produced a comprehensive resource showing the predictive impact of all GLP1R MVs.

These findings establish the GPCR^VP score as a GPCR-specific and reliable predictor for GLP1R MVs that is likely to be of great relevance for clinical decision-making without the time- and cost-consuming generation of experimental data.