Precision Medicine in Metabolic Disease: Can We Predict Who Responds to Which Therapy?

Key Takeaways
Genetically defined conditions provide concrete examples of diagnosis informing drug choice or the need for glucose-lowering treatment.
Routine clinical characteristics can predict relative glycemic response and provide a benchmark for more complex biomarker approaches.
Genetic, metabolic, and protein associations must be interpreted according to their endpoint, comparator, and validation status.
Prospective evaluation must determine whether using a predictor improves care, beyond demonstrating that it predicts outcomes.

What Would It Mean to Predict the Right Therapy?

A clinical trial can establish that a treatment works without answering the question facing an individual patient: would another therapy work better for me? Research in obesity and diabetes increasingly addresses that question through routine clinical measurements, genetic variants, physiological testing, and metabolic or protein signatures. The evidence ranges from randomized comparisons that distinguish drug responses to exploratory investigations of why patients respond differently.

For developers, predicting response could inform patient selection and treatment allocation, both recognized applications in the U.S. Food and Drug Administration’s (FDA) framework for biomarker context of use.¹ The evidence required depends on the decision the biomarker is intended to support.

Interpreting response studies requires distinguishing prediction of an outcome from prediction of treatment benefit. The FDA–National Institutes of Health BEST resource separates prognostic biomarkers, which identify the likelihood of an outcome, from predictive biomarkers, which identify differences in an intervention’s effects.² Better outcomes among biomarker-positive patients receiving a drug may reflect a more favorable prognosis rather than greater benefit from that drug. Establishing predictive value generally requires treatment-versus-control comparisons across biomarker groups

Even a marker that predicts different magnitudes of benefit may leave the preferred treatment unchanged. If the same therapy remains better in both biomarker groups, the marker’s contribution to choosing between treatments may be limited.

The objective must therefore be specific: identify information that improves a defined treatment decision. Predicting glucose lowering, weight loss, tolerability, and prevention of complications are separate tasks, and evidence supporting one should not automatically extend to another.³

When Molecular Diagnosis Changes Treatment

Genetically defined disease provides a concrete starting point. In a randomized crossover study, researchers at Peninsula Medical School and collaborating institutions compared gliclazide and metformin in 36 people with either HNF1A diabetes, a genetically defined form of diabetes, or type 2 diabetes.⁴ Participants with HNF1A diabetes showed substantially greater glucose lowering with gliclazide than metformin and greater gliclazide sensitivity than the matched type 2 diabetes group. The molecular diagnosis identified a difference in relative drug response.

The strength of evidence nevertheless varies across monogenic diabetes subtypes. A systematic review led by researchers at the University of Chicago and the University of Exeter included 146 studies, only six of which were experimental.⁵ It supported sulfonylurea treatment for HNF1A diabetes while identifying substantial uncertainty about optimal treatment in several other subtypes. Small studies, short follow-up, and risk of bias limited the evidence.

The review also supported no glucose-lowering medication for isolated GCK-related hyperglycemia, another genetically defined condition. Studies assessing medication discontinuation did not show deterioration in glycated hemoglobin (HbA1c). Pregnancy was outside the review’s scope and requires separate consideration. Molecular diagnosis can thus inform whether glucose-lowering treatment is needed, as well as which medication to select.

Obesity offers a related historical example. The FDA’s original 2020 approval of setmelanotide linked eligibility to specified genetic forms of obesity involving POMC, PCSK1, or LEPR deficiency.⁶ That approval illustrates selection based on an identified biological defect; it did not establish a genetic prescribing strategy for general polygenic obesity.

Routine Clinical Data Establish a Practical Benchmark

An important benchmark comes from measurements already collected in clinical care. The TriMaster study, led by researchers at the University of Exeter, tested whether body mass index (BMI) and kidney function distinguished glycemic responses to sitagliptin, canagliflozin, and pioglitazone.⁷ Its randomized, double-blind crossover design assigned 525 participants to successive treatment periods, enabling comparisons within individuals.

Average achieved HbA1c was similar across the drugs, but prespecified patient groups revealed differences in relative response. The glycemic advantage of pioglitazone over sitagliptin was greater in participants with BMI above 30 kg/m² than in those at or below that threshold. Kidney function distinguished another comparison: sitagliptin’s relative glycemic advantage over canagliflozin was greater at estimated glomerular filtration rates of 60–90 mL/min/1.73 m² than above 90 mL/min/1.73 m². These findings concern specific drug comparisons and short-term glycemic outcomes, rather than an overall prescribing preference.⁷

Researchers at Exeter and collaborating universities subsequently developed a five-drug-class model using nine routine patient characteristics. It estimated relative glycemic effectiveness for dipeptidyl peptidase-4 (DPP4) inhibitors, glucagon-like peptide-1 (GLP-1) receptor agonists, sodium-glucose cotransporter-2 (SGLT2) inhibitors, sulfonylureas, and thiazolidinediones.⁸

The model underwent validation in separate observational datasets and three randomized trials. In the geographical validation cohort, patients receiving the therapy predicted to provide the greatest glycemic reduction had an observed 12-month HbA1c benefit of 5.3 mmol/mol compared with matched patients receiving another therapy. This was an observational comparison between matched groups, not a randomized benefit from using the model. Its defined task was predicting glycemic effectiveness, rather than ranking treatments across every benefit and risk.

Biological subgrouping must demonstrate its usefulness against that benchmark. In a 2026 analysis, Exeter researchers compared proposed precision strategies in 309 TriMaster participants who received all three medications.⁹ Direct prediction from routine features outperformed the evaluated clinical clustering approach for short-term glycemic optimization. Clinical features also showed stronger associations with differential response than the evaluated genetic scores representing different diabetes-related biological pathways.

Applying a model to another population introduces a further test. An external validation study examined an SGLT2–DPP4 treatment-selection model across major UK self-reported ethnicity groups.¹⁰ Modest recalibration, or adjustment of the model’s predictions, improved performance. After adjustment, predicted differences in glycemic response agreed well with observed differences across the evaluated groups. The study illustrates why a model should be assessed before deployment in a new population.

Genetics and Physiology Offer Emerging Predictors

Genetic predictors become particularly interesting when they connect response with the biology through which a drug acts. Researchers led by the University of Dundee, with Exeter and other collaborators, examined GLP-1 receptor agonist response in 4,571 adults from four prospective observational cohorts and two randomized clinical trials. Their study identified associations involving GLP1R, which encodes the drug receptor, and ARRB1, using six-month HbA1c reduction as the primary outcome.¹¹

The analysis supplies candidate markers and insight into variation in glucose lowering. The next prescribing question is whether genotype improves the choice between treatments beyond information available from clinical characteristics. Assigning medication according to genotype was not the intervention evaluated.

For weight response, researchers at 23andMe reported a study of 27,885 GLP-1-based medication users in Nature in 2026.¹² They identified genetic associations involving self-reported weight loss and gastrointestinal adverse effects and constructed combined genetic and non-genetic response models. The weight-response model explained approximately 25% of variation in percentage BMI loss in training and held-out data, with most explained variation attributable to non-genetic factors. That performance cannot be attributed to genetic testing alone.

The authors also assessed predictions in 642 participants with electronic health record data who had not contributed to model construction. To approximate prediction before treatment, they assigned constant values to unknown dose, duration, and drug-type inputs. This provided a separate prediction assessment, rather than a prospective test of genotype-guided medication selection.

A separate multi-biobank study involving researchers at the University of Helsinki, the Broad Institute, and other institutions provides an important counterpoint. It found no significant associations between GLP-1 receptor agonist-induced weight loss and the evaluated BMI or type 2 diabetes polygenic scores or GLP1R missense variants.¹³ Its conclusions apply to the genetic measures evaluated, rather than establishing that genetic prediction of response is impossible.

Physiological testing offers a complementary approach. Researchers at Mayo Clinic and Phenomix Sciences described a satiation study that combined calories consumed before reaching fullness with genetic modeling.¹⁴ Their model included genetic scores alongside height, weight, age, and sex.

Higher calorie intake before fullness, or a model classification corresponding to that phenotype, was associated with greater weight loss during phentermine–topiramate treatment. Lower measures or corresponding classifications were associated with greater liraglutide response. These observations came from separate trials with different durations, and the liraglutide analysis was post hoc. The authors called for prospective validation using prespecified methods and larger, more diverse populations.

The developmental opportunity is to test whether these classifications help select medication. Prospective trials are needed to determine whether using these classifications to select medication improves weight-loss outcomes.

What Metabolic and Protein Signatures Can Reveal

Metabolic and protein measurements can serve several purposes: identifying baseline differences in treatment benefit, predicting outcomes among treated patients, or characterizing biological changes after therapy begins. Those uses require different interpretations.

Researchers at Beth Israel Deaconess Medical Center, George Washington University, and collaborating institutions investigated baseline metabolic differences in the Diabetes Prevention Program. Their analysis profiled 331 metabolites in 2,015 samples and examined associations with diabetes development across lifestyle intervention, metformin, and placebo groups.¹⁵ Exploratory findings involving phospholipids and adenosine monophosphate suggested differences in preventive treatment effects.

The investigators acknowledged that the analysis was underpowered to definitively establish treatment-effect heterogeneity and called for validation. Its relevance is to preventing diabetes among people at high risk, rather than choosing medication for established disease. The study provides hypotheses about which metabolic characteristics might distinguish preventive benefit.

Another study examined candidate protein predictors during SGLT2 inhibitor treatment. Researchers at National Defense Medical University and Tri-Service General Hospital in Taiwan evaluated 70 patients initiating therapy and identified baseline IL-18/IL-18R1 measurements associated with response.¹⁶ The authors described a single-center observational design without a concurrent control group and characterized the thresholds as provisional. They called for external validation and clinical impact studies before application.

Proteomics can also contribute through mechanistic investigation. Researchers at Novo Nordisk and collaborating academic institutions analyzed baseline and end-of-treatment samples from 1,311 STEP 1 participants and 645 STEP 2 participants.¹⁷ Their study documented circulating protein changes during semaglutide treatment, including changes remaining after accounting for alterations in weight and HbA1c.

These protein changes may help explain semaglutide’s biological effects. Whether baseline protein measurements can identify patients who would benefit more from semaglutide than another therapy remains unproven.

Designing Trials That Validate Treatment Selection

A biomarker program should begin with the decision it intends to support. The FDA describes a biomarker’s context of use through its category and specified application, including enrollment, treatment allocation, dose selection, or response evaluation.¹ Defining the application clarifies what the development program needs to demonstrate.

For treatment selection, validation must assess whether the predictor distinguishes the effects of competing treatments. Prospective implementation addresses a further question: do outcomes improve when clinicians use that prediction to choose therapy? Existing randomized-trial data sets can support the first task even when those trials did not assign care according to the model being evaluated. The five-drug-class study illustrates that distinction between validating predictions and testing a prescribing intervention.^2,8

Prespecification and replication are particularly important when investigators examine many candidate measurements. Mayo Clinic researchers identified multiplicity as a central statistical challenge in biomarker validation and emphasized prespecified hypotheses, randomization, and replication. Testing a defined hypothesis separately from the analyses that generated it helps distinguish validation from continued discovery.¹⁸

Enrollment strategy determines which treatment comparisons remain possible. The FDA’s enrichment guidance explains that studying only marker-positive participants leaves no new clinical evidence about marker-negative patients.¹⁹ Unless earlier findings or a strong mechanistic rationale justify exclusion, the agency encourages including some marker-negative patients in most trials intended to provide primary effectiveness evidence. Uncertainty about the cutoff and response outside the selected group increases the importance of that assessment.

The endpoint must also remain explicit. A systematic review led by researchers at Exeter and Dundee found substantial methodological limitations in evidence on treatment-effect heterogeneity for GLP-1 receptor agonists and SGLT2 inhibitors.³ Observational studies suggested relationships between kidney function and glycemic response to SGLT2 inhibitors, and between reduced insulin secretion and glycemic response to GLP-1 receptor agonists. Among the cardiovascular and renal analyses included in the review, the investigators found minimal clinically relevant treatment-effect heterogeneity.

A predictor of glucose lowering should consequently remain a predictor of that outcome unless additional evidence establishes another use. Selecting a drug to reduce HbA1c and selecting one to prevent a kidney complication require their own supporting comparisons.

Making Prediction Useful in Clinical Practice

Dependable measurement is a prerequisite for applying a biological predictor. The FDA’s April 2026 guidance on bioanalytical method validation for biomarkers addresses assays measuring biomarker concentrations in biological matrices.20 It recommends a fit-for-purpose approach covering accuracy, precision, sensitivity, selectivity, reproducibility, stability, and sample collection, handling, and storage.

The guidance distinguishes internal development decisions from regulatory uses supporting approval or labeling, for which fuller validation is recommended. Its scope concerns concentration assays, rather than providing a universal framework for genetic testing or computational models. An assay used to support a treatment decision must generate sufficiently reliable measurements for that purpose.

Once reliable measurements and predictions enter care, the complete intervention becomes the subject of evaluation. Researchers at the Chinese University of Hong Kong reported the Precision Medicine to Redefine Insulin Secretion and Monogenic Diabetes (PRISM) trial in 2026.21 The three-year randomized trial enrolled 884 Chinese adults with young-onset diabetes. Its intervention combined autoantibodies, C-peptide, genetic information, intensified algorithm-based treatment, and multidisciplinary follow-up.

More intervention participants achieved three or more treatment targets than those receiving usual clinic-based care: 23.8% versus 12.2%. However, the trial did not demonstrate a reduction in its primary complication outcome, which occurred in 26.3% versus 28.2%. PRISM evaluated the complete care program; it did not isolate the contribution of an individual biomarker. Its improved target attainment and primary outcome result therefore need to be considered together.

Precision medicine in metabolic disease already has treatment-relevant applications in selected molecular diagnoses, and routine clinical models have demonstrated the ability to predict relative glycemic response. Genetic, physiological, metabolic, and protein approaches offer additional candidates at different stages of development. Broader use for medication selection in common obesity and diabetes requires evidence that the added information improves a specified choice and that acting on it produces better outcomes. That is the practical standard against which new predictive tests should be judged.

References

1. “Context of Use.” U.S. Food and Drug Administration. Accessed 02 Nov. 2026.

2. FDA-NIH Biomarker Working Group. “BEST (Biomarkers, EndpointS, and other Tools) Resource.” Food and Drug Administration (US) and National Institutes of Health (US). 2016–.

3. Young, Katherine G, et al. “Treatment effect heterogeneity following type 2 diabetes treatment with GLP1-receptor agonists and SGLT2-inhibitors: a systematic review.” Communications Medicine. 3: 131 (2023).

4. Pearson, Ewan R, et al. “Genetic cause of hyperglycaemia and response to treatment in diabetes.” The Lancet. 362: 1275–1281 (2003).

5. Naylor, Rochelle N, et al. “Precision treatment of beta-cell monogenic diabetes: a systematic review.” Communications Medicine. 4: 145 (2024).

6. “FDA approves first treatment for weight management for people with certain rare genetic conditions.” U.S. Food and Drug Administration. 27 Nov. 2020.

7. Shields, Beverley M, et al. “Patient stratification for determining optimal second-line and third-line therapy for type 2 diabetes: the TriMaster study.” Nature Medicine. 29: 376–383 (2023).

8. Dennis, John M, et al. “A five-drug class model using routinely available clinical features to optimise prescribing in type 2 diabetes: a prediction model development and validation study.” The Lancet. 405: 701–714 (2025).

9. Güdemann, Laura M, et al. “To Cluster or Not to Cluster? A Comparison of Approaches to Targeting Type 2 Diabetes Glucose-Lowering Therapy in the TriMaster Crossover Trial.” Diabetes Care. 49: 1458–1466 (2026).

10. Güdemann, Laura M, et al. “Validation of an algorithm for selection of SGLT2 and DPP4 inhibitor therapies in people with type 2 diabetes across major UK ethnicity groups: a retrospective cohort study.” The Lancet Regional Health – Europe. 61: 101547 (2026).

11. Dawed, Adem Y, et al. “Pharmacogenomics of GLP-1 receptor agonists: a genome-wide analysis of observational data and large randomised controlled trials.” The Lancet Diabetes & Endocrinology. 11: 33–41 (2023).

12. Su, Qiaojuan Jane, et al. “Genetic predictors of GLP1 receptor agonist weight loss and side effects.” Nature. 653: 770–775 (2026).

13. German, Jakob, et al. “Association between plausible genetic factors and weight loss from GLP1-RA and bariatric surgery.” Nature Medicine. 31: 2269–2276 (2025).

14. Cifuentes, Lizeth, et al. “Genetic and physiological insights into satiation variability predict responses to obesity treatment.” Cell Metabolism. 37: 1655–1666.e5 (2025).