A receptor-specific variant prediction model improved prediction of the functional effects of glucagon-like peptide-1 receptor variants, reducing false-positive predictions by approximately 78% compared with a general-purpose predictor across two validation sets, according to a study published in Diabetologia.  

Only a fraction of known glucagon-like peptide-1 receptor (GLP1R) missense variants have been experimentally characterized, the researchers wrote. They added that existing variant-effect predictors are not optimized for membrane proteins such as G protein-coupled receptors (GPCRs).  

To address this gap, the researchers developed a GPCR variant impact predictor (GPCRVP) using a training set of 58 experimentally characterized GLP1R variants. The model combined eight GPCR-specific structural and dynamic features with predictions from AlphaMissense, a deep-learning model, and REVEL, a random-forest ensemble predictor. The experimental functional-impact score incorporated receptor surface expression and two measures of cAMP signaling: potency (EC50) and efficacy (Emax).

The GPCR-specific features incorporated membrane context, allosteric communication, changes in receptor connectivity, ligand and G protein interactions, location-specific amino acid substitution favorability, BLOSUM62 scores, charge changes, and helix-breaking propensity.

GPCR-specific model improves prediction 

Among standalone predictors evaluated in the 58-variant training set, AlphaMissense showed the strongest correlation with experimentally measured functional effects (Pearson’s r=0.754; root mean square error [RMSE], 0.222), followed by REVEL (r=0.722; RMSE, 0.199). The best-performing integrated model combined the GPCR-specific features with both AlphaMissense and REVEL. The integrated GPCRVP model showed the strongest correlation with experimental measurements (r=0.858) and the lowest RMSE (0.145). 

The model was then tested on 50 mutations not used to train the model: an internal validation set of 15 newly characterized mutations and an external validation set of 35 mutations from an independent study. The internal set was selected based on GPCRVP scores to span low- to high-impact changes and included several mutations for which AlphaMissense and GPCRVP produced substantially different predictions.

Across the combined validation sets, AlphaMissense produced nine false-positive predictions and one false negative using its established classification thresholds. In comparison, GPCRVP produced two false positives, an approximately 78% reduction, and no false negatives. Prediction error was also lower with GPCRVP than AlphaMissense, with RMSE values of 0.185 and 0.379, respectively.  

Extending predictions to human variants 

The researchers subsequently constructed structural models for 174 known human GLP1R missense variants located in experimentally resolved receptor regions and not included in the training or validation sets, then applied GPCRVP to these variants.

Rare variants tended to receive higher predicted functional-impact scores, whereas more common variants were generally predicted to have wild-type-like function. Variants predicted to have greater impact were concentrated in the core of the transmembrane region and upper portion of the extracellular domain.  

By extrapolating false-positive ratios observed during model development and validation, the researchers projected false-positive rates of 8% for AlphaMissense and 3% for GPCRVP across GLP1R variants in experimentally resolved receptor regions at the AlphaMissense threshold of 0.564. At a threshold of 0.8, the projected rates were 4% and 0%, respectively.  

In a separate small case study of eight randomly selected missense variants in the metabolically relevant melanocortin 4 receptor, AlphaMissense misclassified three variants, whereas GPCRVP correctly classified all eight based on reported experimental data. However, the researchers cautioned that the small sample requires further validation.  

Limitations and potential applications 

The researchers noted several limitations. GPCRVP was designed to predict variant effects on receptor function rather than pathogenicity, and the researchers cautioned that computational predictions of receptor function may not capture complex in vivo effects, including treatment response. The model also cannot assess variants in structurally unresolved receptor regions. Because it was developed primarily using GLP1R variants, broader benchmarking across diverse G protein-coupled receptor families is needed to establish its generalizability. 

The researchers proposed that predicting the functional impact of GLP1R variants could aid interpretation of genetic findings in patients with atypical metabolic phenotypes, including unexpected responses to GLP-1–based therapies, and could support patient stratification and pharmacogenomic research. These potential applications were not evaluated clinically in the study.  

“By integrating structural, dynamic, and membrane-aware features that reflect core aspects of GPCR biology, the model improves the accuracy, interpretability, and clinical relevance for GLP1R variant assessment,” the researchers wrote. 

Tomasz M. Stepniewski, PhD, reported being an employee of InterAx Biotech. The remaining authors declared no conflicts of interest.