Machine learning-guided optimization of triple agonist peptide therapeutics for metabolic disease.

Introduction: Multi-target peptide therapeutics targeting glucagon receptor (GCGR), glucagon-like peptide-1 receptor (GLP1R), and glucose-dependent insulinotropic polypeptide receptor (GIPR) represent a promising approach for treating diabetes and obesity. Triple agonist peptides demonstrate promising therapeutic potential compared to single-target approaches, yet rational design remains computationally challenging due to complex sequence-structure activity relationships. Existing methods, primarily based on convolutional neural networks, impose limitations including fixed sequence lengths and inadequate representation of molecular topology. Graph Attention Networks (GAT) offer advantages in capturing molecular structures and variable-length peptide sequences while providing interpretable insights into receptor-specific binding determinants. Methods: A dataset of 234 peptide sequences with experimentally determined binding affinities was compiled from multiple sources. Peptides were represented as molecular graphs with seven-dimensional node features encoding physicochemical properties and positional information. The GAT architecture employed a shared encoder with task-specific prediction heads, implementing transfer learning to address limited GIPR training data. Performance was evaluated using 5-fold cross-validation and independent validation on 24 literature-derived sequences. A genetic algorithm framework was developed for peptide sequence optimization, incorporating multi objective fitness evaluation based on predicted binding affinity, biological plausibility, and sequence novelty. Results: Cross-validation demonstrated robust GAT performance across all receptors, with GCGR achieving high accuracy (AUC ROC: 0.915 ± 0.050), followed by GLP1R (AUC-ROC: 0.853 ± 0.059), and GIPR showing acceptable performance despite limited data (AUC-ROC: 0.907 ± 0.083). Comparative analysis revealed receptor-specific advantages: GAT significantly outperformed CNN for GCGR prediction (RMSE: 0.942 vs. 1.209, p = 0.0013), while CNN maintained superior GLP1R performance (RMSE: 0.552 vs. 0.723). Genetic algorithm optimization measurable improvement over baseline, with 4.0% fitness Enhancement and generation of 20 candidates exhibiting mean binding probabilities exceeding 0.5 across all targets. The GAT-based framework provides a computational approach in computational peptide design, demonstrating receptor-specific advantages and robust optimization capabilities. Conclusion: Genetic algorithm optimization enables systematic exploration of sequence space within existing agonist scaffolds while maintaining biological constraints. This approach provides a rational framework for prioritizing experimental validation efforts in triple agonist development.