Abstract
The discovery of peptide therapeutics requires the simultaneous optimization of multiple, often conflicting, parameters including potency, selectivity, and biophysical stability. We describe streaMLine, an AI-driven discovery platform that utilizes a closed-loop cycle of high-throughput screening and machine learning. By applying this platform to a biophysically robust Secretin scaffold, we successfully engineered GUB021796, a highly potent and selective GLP-1 receptor (GLP-1R) agonist. GUB021796 demonstrates a 71-hour half-life in mini-pigs and significant weight-loss efficacy in mouse models, highlighting the power of ML-guided peptide engineering.
Introduction
Peptide drug discovery for metabolic diseases, particularly GLP-1 receptor agonists, faces the challenge of balancing therapeutic potency with manufacturing feasibility and pharmacokinetic longevity. Traditional "trial and error" methods struggle to optimize these parameters in parallel. This study addresses these hurdles by utilizing the streaMLine platform to transform Secretin—a peptide with superior biophysical stability but low native GLP-1R activity—into a potent therapeutic candidate.
Methodology: The streaMLine Platform
The streaMLine platform functions as an iterative, closed-loop discovery cycle designed to generate high-quality, project-specific data.
The process begins with library generation and rapid solid-phase peptide synthesis. These peptides undergo high-throughput multi-parameter screening, encompassing functional assays, biophysical characterization (e.g., solubility, fibrillation), and metabolic stability. The resulting data trains machine learning models to predict the properties of unsynthesized sequences. To ensure interpretability, the team employed SHAP (SHapley Additive exPlanations) values to visualize how specific amino acid substitutions at each position contribute to the desired profile.
Results and Discussion
Deep Mutational Scanning for Selectivity
A critical component of the engineering process involved "positional mapping" or deep mutational scanning of the peptide backbone. This allowed researchers to identify specific residues that govern the selectivity between the GLP-1 and Secretin receptors.
As shown in the mutational analysis of position P18, different amino acid substitutions (such as Aib, Nle, or L) significantly impact the balance between hGLP1R and hSCTR activation. By leveraging these insights, the researchers were able to "dial down" Secretin receptor activity while maximizing GLP-1R potency.
Characterization of GUB021796
The optimization process culminated in the identification of the development candidate GUB021796. This molecule achieved a GLP-1R EC50 of 0.015 nM, making it highly potent and comparable to native GLP-1, while maintaining high selectivity (EC50 of 190 nM at the Secretin receptor). Biophysically, GUB021796 exhibited no fibrillation and high solubility (>10 mg/mL), addressing common manufacturing and delivery hurdles. In pharmacokinetic studies, the candidate showed a half-life of 71 hours in mini-pigs, supporting the potential for once-weekly human dosing. In vivo efficacy was confirmed in mouse models, where GUB021796 produced significant, dose-dependent body weight loss over 24 days.
Conclusion
The streaMLine platform demonstrates that machine learning can effectively bridge the gap between biophysically robust but inactive scaffolds and potent therapeutic agonists. By providing a "map" of the peptide sequence space, ML allows for the parallel optimization of potency, selectivity, and stability. Future work will focus on applying this target-agnostic platform to other challenging peptide GPCR targets.
Summary updated