Can AI improve GLP-1 and weight-loss treatment?
A new analysis of research from China and South Korea: predicting treatment response and designing a different drug candidate are promising ideas, but neither establishes that using AI improves weight-loss outcomes.
The short version
- What happened
- Chinese peer-reviewed work explores treatment selection for blood-sugar control; that endpoint must not be relabelled weight loss.
- Who it affects
- Readers can distinguish research aimed at selecting treatment from research aimed at inventing a medicine.
- What remains unproven
- The cited treatment-selection paper does not test weight-loss benefit from AI-guided care.
Keep in mind
The cited treatment-selection paper does not test weight-loss benefit from AI-guided care.
Hanmi's poster and pipeline are commercially interested primary materials, with preclinical claims and company-reported trial status.
The review's search ends in June 2026, leaving later developments outside its assessment.
This feature compares selected examples; it is not an exhaustive review or a treatment recommendation.
Multi-source analysis · 4 linked sources. Primary analysed source date: 17 February 2026. Earliest linked supporting source: 4 November 2025. Latest linked supporting source: 3 September 2026.
Read the full report · 6 min The evidence, methods, implications and limitations
Two useful stories, with different kinds of evidence
AI and metabolic medicines meet at two different questions: can a model help choose treatment, and can computation help create a new molecule? This feature compares examples originating in China and South Korea. It is analysis published on 5 October, not a claim that a medicine has just been approved. The Chinese papers appeared in February and September 2026; the Korean programme's linked discovery poster dates to November 2025.
Our editorial assessment is that the distinction matters more than a headline about an 'AI weight-loss breakthrough'. A response model is a decision-support proposal. A designed molecule is a drug-development proposal. Each needs evidence appropriate to its intended use, and neither should inherit the clinical reputation of an established medicine simply because the same story mentions GLP-1.[1][2][3][4]
China: an AI model asks which diabetes treatment may work better
The Shanghai-led TiP DecScore study used gradient-boosted decision trees and 15 clinical features to compare GLP-1 receptor agonists with SGLT-2 inhibitors. It developed the model in 24,322 patients across 254 Chinese centres and tested it in a separate 1,459-patient Shanghai dataset. Outcomes were HbA1c control and levels at six and twelve months. Reported AUROCs were 0.71–0.78; treatment concordance was associated with better control. This was retrospective analysis, not a randomised trial of AI-guided care or a weight-loss efficacy test.
A discrimination score describes how well a model separates outcomes in its evaluation data. It is not the percentage of patients who will benefit if clinicians follow it. More generally, associations between a recommendation and an outcome can reflect differences in who receives treatment. Our interpretation is that this is a candidate for prospective evaluation, with the endpoint kept explicit, rather than a ready-made answer for selecting an obesity medicine.[1]
South Korea: a different candidate, rather than a new GLP-1
Hanmi describes HM17321 as a long-acting urocortin-2 analogue targeting CRF2, rather than a GLP-1 drug. Its discovery poster combines structural modelling, a protein-activity prediction workflow called HARP-pSAR, laboratory assays and animal experiments. The company describes body-composition effects in nonclinical models. Its current pipeline lists a US Phase 1 study and links registration NCT07219589. That status is attributed to Hanmi here; we could not independently verify the registry record during this review. The cited materials do not establish human weight-loss efficacy or regulatory approval.
This creates a worthwhile reporting question: what did the computational component contribute beyond the rest of the discovery process? A convincing answer would compare candidate selection, experimental work and subsequent outcomes against a defined alternative. Our assessment is that a promising programme can justify further investigation without proving either that AI caused the progress or that the resulting molecule will become a useful treatment.[3][4]
A September review provides context, with a June cutoff
The Chengdu-based Frontiers review searched literature through 1 June 2026. Within its eligible evidence, it found no externally validated, prospectively implemented AI prescribing system demonstrating improved patient outcomes. Drug discovery was largely preclinical or early translational. This is a structured narrative review, not a pooled estimate of AI's weight-loss effect, and its cutoff cannot settle what happened afterwards.
Our reading is that the review supplies a useful checklist for the next study: distinguish model validation from implementation, implementation from benefit, and an entire care package from the contribution of AI. If a programme combines clinician contact, coaching, medication and a digital tool, improvement alone does not identify which component helped. A comparison designed around that question would be more informative than an undifferentiated success claim.[2]
The next test should follow the intended clinical claim
For treatment selection, we would look for a prospectively specified comparison of usual care with model-supported care, measuring the outcome the tool actually seeks to improve. Evaluation should include whether clinicians use recommendations, why they override them, and whether uncertainty is communicated. A tool that performs well on archived records could still fail to change decisions or could direct resources toward patients who were already likely to do well.
For a new drug candidate, we would look for transparent human safety and efficacy reports, suitable comparison groups and follow-up sufficient for the claims being made. If preservation of lean mass becomes a proposed advantage, the evaluation should examine function and how body composition was measured. Our standard would be the same regardless of whether the molecule emerged from an AI workflow or conventional research.[1][2][3][4]
What this could mean for people and health systems
The opportunity is to make care more responsive to an individual's needs. But our assessment would include affordability, access to follow-up and the burden of collecting inputs, alongside a predictive score. A recommendation that cannot be acted on because a medicine is unavailable is not equivalent to improved care. Nor should an uncertain model replace a discussion about the outcomes a person values.
For cross-country use, the next question is whether the intended service resembles the setting in which the evidence was developed. We would want validation in the receiving population and reporting on calibration, missing data and differences between patient groups. Research originating in East Asia deserves attention on its own terms; geographical origin does not make a model universally transferable or determine where a medicine will be tested or licensed.[1][2]