Abstract: In recent years, GLP-1 receptor agonists (GLP-1RAs), led by semaglutide, have revolutionized the treatment of obesity and type 2 diabetes worldwide. However, even with a once-weekly dosing regimen, there is still room to improve long-term patient adherence. A core bottleneck in drug development has always been how to further extend a drug's half-life to achieve less frequent administration without compromising efficacy. A recent study published in Advanced Science leverages AI-driven de novo design to successfully develop novel GLP-1RA molecules with half-lives far exceeding that of semaglutide. This work provides a clear technological path toward a "monthly injection" and signals a paradigm shift in how AI is reshaping peptide drug discovery [1].

The Evolution of GLP-1 Drugs: An Unending Quest for Longevity

The history of GLP-1 drug development is essentially a story of battling the body's rapid degradation mechanisms. Native GLP-1 has a half-life of only a few minutes in vivo, making it unsuitable as a therapeutic. From the 2005 approval of the first GLP-1RA, exenatide, to today's blockbuster drugs, technological innovation has consistently centered on the goal of "long-acting" formulations [2].

• First Generation: Mimicking Nature, Initial Success. Exenatide, derived from the Gila monster's Exendin-4, shares structural similarity with human GLP-1 but is more resistant to DPP-4 enzyme degradation. This extended its half-life to the order of hours, but still required twice-daily injections [3].
• Second Generation: Chemical Modification Achieves Weekly Dosing. Liraglutide and semaglutide represent this leap, incorporating a fatty acid chain that enables binding to serum albumin. By leveraging albumin's long half-life (approx. 19 days), these drugs evade rapid renal clearance, successfully reducing dosing frequency to once daily or even once weekly [4]. Semaglutide further enhanced stability and affinity through amino acid substitutions, setting the benchmark for weekly injectables.
• The New Challenge: Breaking the "Weekly" Barrier. While weekly formulations are a monumental success, the pace of drug development has not slowed. The advent of dual- and triple-target agonists (e.g., tirzepatide, retatrutide) has achieved breakthroughs in efficacy [5]. However, in the dimension of "duration of action," the systematic and predictable design of "ultra-long-acting" molecules that surpass existing albumin-binding strategies remains a formidable challenge. Traditional, trial-and-error sequence modification methods are inefficient and struggle to escape the confines of existing molecular scaffolds.

AI-Driven De Novo Design for Precision-Engineered, Ultra-Long-Acting Molecules

The study in Advanced Science offers a novel solution to this bottleneck: instead of merely fine-tuning existing molecules, it uses generative AI to design entirely new peptide sequences with superior properties from the ground up [1].

2.1 Design Strategy: Fixed Backbone, AI-Generated Variables

The research team employed a strategy of "fixing the critical backbone, letting AI explore the rest." They began by analyzing the crystal structure of semaglutide bound to the GLP-1 receptor (PDB 7KI0) to identify 13 conserved amino acid residues essential for receptor binding. These sites, like the load-bearing columns of a building, were kept constant.

Next, they "opened up" the remaining 18 non-critical positions to a deep learning model called ProteinMPNN. ProteinMPNN is an advanced protein sequence design tool that can, given a 3D backbone structure, predict amino acid sequences likely to fold into that structure stably [6]. The model generated 10,000 novel sequences in a single run.

2.2 Computational Funnel: From 10,000 to 60

Faced with a vast number of candidates, the team established an efficient computational screening pipeline:

1. Initial Filtering: First, redundant sequences were removed, leaving 1,443 unique candidates. Then, sequences containing known protease cleavage sites were removed. Physicochemical properties like net charge and hydrophobicity were used to filter out potentially unstable molecules.
2. Structural Validation: AlphaFold2 was used to predict the 3D structure of each candidate in complex with the GLP-1 receptor. Metrics such as pLDDT (prediction confidence) and RMSD (deviation from the original structure) ensured that the new sequences would fold correctly and bind the receptor, retaining only the most structurally reliable candidates.
3. Affinity Assessment: Molecular dynamics (MD) simulations were performed to calculate the binding free energy between each sequence and the receptor, further refining the list to the top 60 candidates for wet-lab validation.

This purely computational workflow accomplished in weeks what would traditionally take months or even years, dramatically accelerating the research and development process.

2.3 Experimental Validation: From In Vitro Activity to In Vivo Triumph

Of the 60 candidates, six demonstrated good receptor affinity in in vitro SPR binding assays. Among them, three molecules—D13, D41, and D44—were particularly outstanding:

• In Vitro Activity: D41 and D44 exhibited EC50 values (half-maximal effective concentration) superior to semaglutide, indicating more potent receptor agonism.
• In Vivo Pharmacokinetics (PK): The true breakthrough was revealed in rat PK studies. When attached to the same C20 fatty acid chain (semaglutide uses C18), D41 achieved a half-life of 23.16 hours, 2.83 times that of semaglutide (8.17 hours), while D13 reached a 2.43-fold increase. This exponential improvement confirmed that the novel, AI-designed sequence played a decisive role in extending the half-life.
• Animal Model Efficacy: In diabetic (db/db) and obese (DIO) mouse models, D13 and D41 not only demonstrated more sustained glucose-lowering effects than semaglutide but also achieved nearly identical weight loss (19.25% for D13 vs. 21.54% for semaglutide), while significantly improving lipid profiles and other metabolic markers.

Ushering in a New Paradigm for Peptide Drug Design

The significance of this research extends far beyond the creation of a few promising GLP-1RA candidates. It showcases a new R&D paradigm: the accelerated realization of the Design-Build-Test-Learn (DBTL) cycle.

• From "Experience-Driven" to "Data-Driven": Traditional peptide drug optimization relies heavily on expert chemical intuition and extensive trial and error. This study demonstrates how AI models can efficiently and predictably explore a vast sequence space, transitioning drug discovery into a more data-driven era.
• Methodological Universality: The strategy of "fixed functional backbone + AI-generated variable regions + computational screening" is highly generalizable. It can be applied not only to optimize other peptide drugs (e.g., insulin, growth hormone) but also for engineering more complex proteins like antibodies and enzymes.
• Accelerating the DBTL Cycle: Challenges and Opportunities: While AI has shown incredible efficiency in the "Design" phase, the large-scale "Build" and "Test" phases remain a bottleneck. The ability to rapidly synthesize and screen thousands of candidates is crucial to the overall speed of the DBTL cycle. This presents a significant opportunity for high-throughput gene construction and expression screening platforms. For instance, technologies like Ailurus Bio's Ailurus vec®, a self-selecting expression vector system, could dramatically accelerate this process by linking target expression to host survival, enabling parallel screening of massive construct libraries in a single experiment to quickly identify optimal expression solutions.

Conclusion

From exenatide to semaglutide, and now to AI-designed ultra-long-acting molecules, the evolution of GLP-1 drugs clearly illustrates the trajectory of biomedical technology. This paper in Advanced Science not only brings the vision of a "monthly injection" for diabetes and obesity treatment closer to reality but, more importantly, reveals how AI can serve as a powerful engine to transform drug R&D from a labor-intensive process to an efficient, precise, and "intelligent creation" model. In the future, with the deep integration of AI algorithms, high-throughput experimental technologies, and automated platforms, we can confidently expect that more "impossible" drug molecules will be designed and created, truly changing the future of human health.

References

1. Wei, T., et al. (2025). AI-Driven De Novo Design of Ultra Long-Acting GLP-1 Receptor Agonists. Advanced Science. DOI: 10.1002/advs.202507044.
2. Nauck, M. A., & Meier, J. J. (2018). The incretin effect in healthy individuals and those with type 2 diabetes: physiology, pathophysiology, and response to therapeutic interventions. The Lancet Diabetes & Endocrinology, 6(3), 232-246.
3. Drucker, D. J. (2018). Mechanisms of Action and Therapeutic Application of Glucagon-like Peptide-1. Cell Metabolism, 27(4), 740-756.
4. Lau, J., Bloch, P., Schäffer, L., et al. (2015). Discovery of the Once-Weekly Glucagon-Like Peptide-1 (GLP-1) Analogue Semaglutide. Journal of Medicinal Chemistry, 58(18), 7370-7380.
5. Jastreboff, A. M., et al. (2022). Tirzepatide Once Weekly for the Treatment of Obesity. New England Journal of Medicine, 387(3), 205-216.
6. Dauparas, J., et al. (2022). Robust deep learning–based protein sequence design using ProteinMPNN. Science, 378(6615), 49-56.