Higher semaglutide dose is associated with lower neuropsychiatric event incidence independent of weight loss

While GLP-1 receptor agonists (GLP-1RAs) were originally developed to improve glycemic control in type 2 diabetes, there is growing evidence from randomized clinical trials that these therapies also provide benefits in obesity, cardiovascular disease, chronic kidney disease, obesity-related heart failure with preserved ejection fraction, metabolic dysfunction-associated liver disease, and sleep apnea^{1,2,3,4,5,6,7,8,9}. These successes raise the question of which treatment-related benefits are due to direct pharmacology and which are secondary to weight loss or other systemic metabolic changes?

The answer is unlikely to come from whole-body summary measures alone. Human and translational studies have established that the GLP-1 receptor plays important roles in hypothalamic homeostatic circuits, hindbrain interoceptive pathways, and reward-linked striatal systems, while peripheral tissues such as heart, vasculature, gut, and endocrine organs may add additional on-target or context-dependent effects^{10,11,12,13,14,15,16,17}. At the same time, tissue-level localization of the GLP-1 receptor transcript and protein outside canonical pancreatic compartments remains sensitive to assay choice and bulk-tissue composition, making direct translational inference challenging.

To help deconvolute complex pharmacology of GLP-1RAs, particularly in the context of cognitive and neuropsychiatric outcomes, this work leverages electronic health record (EHR) data from a federated network of 489,785 semaglutide-treated patients^18,19,20. We assessed 24 neuropsychiatric outcomes in a two-year period following semaglutide exposure to identify phenotypes associated with GLP-1 receptor agonism, which could be related to treatment-driven broad metabolic improvements, direct pharmacologic activity in the central nervous system (CNS), or other indirect mechanisms. We then integrated these findings with publicly available transcriptomic datasets to provide descriptive molecular context and generate hypotheses for future mechanistic investigation.

Semaglutide is associated with reduced rates of affective, substance-use, vascular, and neurocognitive outcomes

From a cohort of 489,785 patients with at least one semaglutide prescription, a neuropsychiatric analysis cohort of 63,215 patients with baseline neurological conditions was assembled (see Methods). Of these, 56,859 could be propensity matched 1:1 to metformin-treated patients, 4757 to DPP-4 inhibitor-treated patients, and 30,680 to SGLT2 inhibitor-treated patients. Across 24 incident neuropsychiatric outcomes, we evaluated approximate 2-year event risk after treatment initiation.

Semaglutide was associated with lower two-year event risk than metformin for dementia/degenerative CNS disease (0.95% versus 1.96%; P < 0.001), cognitive symptoms (3.1 versus 4.3%; P < 0.001), substance-related neuropsychiatric disorders (3.8 versus 6.1%; P < 0.001), psychotic disorders (0.31 versus 0.93%; P < 0.001), speech/language symptoms (0.63 versus 1.21%; P < 0.001), cerebrovascular disease (2.6 versus 3.7%; P < 0.001), polyneuropathies (3.5 versus 4.5%; P < 0.001), mood disorders (10.5 versus 13.2%; P < 0.001), personality/impulse-control disorders (0.53 versus 1.06%; P < 0.001), perceptual symptoms (0.24 versus 0.60%; P < 0.001), and neurocognitive disorders (1.0 versus 1.7%; P < 0.001) (Fig. 1A). By contrast, metformin was associated with a lower rate of incident neuromuscular disease (1.1 versus 0.7%; P < 0.001). There was no significant difference in the risk of the negative control outcomes (eczema, kidney stones, rotator cuff disorders, hernia, tinea corporis) between semaglutide and metformin (Fig. S1).

Subgroup-specific comparator analyses were then performed within individual baseline neurological condition strata (see Methods, Tables 1–3, and Tables S1–4). The largest associations with semaglutide were observed in patients with baseline neurocognitive disorders, where semaglutide was associated with a lower subsequent incidence of mood disorders. (13.7 versus 18.0%; P = 0.017). The strongest metformin-favored effects included lower rates of incident seizure and paroxysmal neurologic disorders among patients with baseline neurocognitive disorders (33.2 versus 24.3%; P = 0.015), as well as lower rates of incident neuromuscular disease among those with baseline polyneuropathy/peripheral nervous system disorders (3.0 versus 1.1%; P < 0.001) and Parkinson’s/movement disorders (2.2 versus 0.8%; P < 0.001). Broadly similar but mixed patterns were observed in the corresponding comparisons with DPP-4 inhibitors (Fig. 1B) and SGLT-2 inhibitors (Fig. 1C), with semaglutide generally associated with a lower incidence of several neuropsychiatric outcomes.

Maximum attained semaglutide dose was associated with greater maximum weight loss through the pre-landmark period

Maximum attained semaglutide dose during the first 2 years after treatment initiation, defined as the pre-landmark period for subsequent analyses, was associated with greater maximum weight loss achieved within that same interval (Fig. 2). Patients whose maximum dose remained at 0.25 mg or 0.5 mg showed the least weight reduction, whereas those reaching 1.7, 2.0, and especially 2.4 mg had progressively greater negative body weight change distributions. The maximum achieved body weight change increased by about 3.01% for each 1 mg increase in semaglutide dose.

Weight loss-stratified analyses show little association with most neuropsychiatric outcomes

When patients were stratified by maximum achieved weight loss during the pre-landmark period, most post-landmark neuropsychiatric and neurological outcomes did not differ significantly across weight loss categories (Fig. 3 and Fig. S2). Baseline characteristics and semaglutide prescription patterns across weight loss strata are summarized in Table 1. Global weight loss-strata tests were non-significant for substance-related neuropsychiatric disorders (P = 0.15), mood/affective disorders (P = 0.96), and hereditary/systemic CNS atrophies (P = 0.99), with cumulative event curves showing no significant separation across the <5, 5–10, 10–15, 15–20, and ≥20% weight loss groups. These findings suggest that, within this landmark framework, achieved weight loss was not strongly associated with the subsequent incidence of neuropsychiatric conditions. The negative control outcomes were also not significantly associated with the magnitude of weight change (Fig. S3).

Reduced rates of neuropsychiatric outcomes in patients achieving higher semaglutide doses

In the landmark analysis (Fig. 4 and Fig. S4), higher attained semaglutide dose (≥1.7 mg) during the pre-landmark period was associated with lower post-landmark event rates than lower dose regimens (0.25–1.0 mg). Baseline characteristics and pre- versus post-landmark semaglutide prescription patterns across these dose strata are summarized in Table 2. Higher attained dose was associated with lower post-landmark event rates for substance-related neuropsychiatric disorders (RR 0.71, 95% CI 0.62-0.81; P < 0.001), mood/affective disorders (RR 0.82, 95% CI 0.75–0.89; P < 0.001, hereditary/systemic CNS atrophies (RR 0.59, 95% CI 0.45–0.78; P < 0.001), anxiety/trauma/OCD/somatoform disorders (RR 0.88, 95% CI 0.81–0.94; P < 0.001), neuromuscular junction or muscle disease (RR 0.72, 95% CI 0.55–0.93; P = 0.013), eating/sleep/behavioral syndromes (RR 0.87, 95% CI 0.76–0.99; P = 0.037), personality/impulse-control disorders (RR 0.68, 95% CI 0.48–0.96; P = 0.028), and neurologic signs and symptoms (RR 0.87, 95% CI 0.76–0.99; P = 0.037). On the other hand, a higher attained dose was not significantly associated with any of the negative control outcomes (Fig. S5).

Weight loss-dominant outcomes separate from the dose-dominant cluster

To distinguish dose-associated from weight loss-associated patterns, post-landmark neurological and neuropsychiatric outcomes were next evaluated along two complementary axes: the strength of association with attained dose and the strength of association with achieved weight loss. Under this framework, substance-related neuropsychiatric disorders, mood/affective disorders, hereditary/systemic CNS atrophies, anxiety/trauma/OCD/somatoform disorders, and residual CNS and neurologic disorders were more strongly associated with dose, whereas cognitive symptoms and speech/language symptoms showed stronger associations with weight change (Fig. 5 and Table 3). This association with weight loss was particularly strong for cognitive symptoms (global P < 0.001), with approximate post-landmark event rates of 2.1% in the <5% weight loss group compared to 5.5% in the 15–20% group and 4.7% in the ≥20% group (Fig. S6). Speech/language symptoms showed a similar but less prominent pattern (global P = 0.003). Together, these findings suggest that at least some of the neuropsychiatric associations observed with semaglutide are more closely correlated with maximum prescribed dose than with weight reduction.

A descriptive atlas of GLP1R transcript expression across human tissues and CNS cell populations

As expected, bulk RNA-sequencing data across human tissues highlighted the pancreas as the highest GLP1R-expressing tissue, including endocrine and exocrine components (Fig. 6A and Table S3). However, several non-pancreatic sites with median GLP1R expression of at least 1 transcript per million (TPM) were noted, including in the heart (atrial appendage), CNS (hypothalamus, caudate), and the peripheral nervous system (tibial nerve) (Fig. 6B). Among the CNS sites of possible expression, the highest expression was seen in the caudate nucleus of the basal ganglia (median 1.50 TPM, n = 300) and hypothalamus (median 1.21 TPM, n = 257) (Fig. 6B), although the observed transcript abundance remained low and should be interpreted cautiously given the known challenges of detecting low-abundance GLP1R transcripts in bulk RNA-sequencing datasets.^{21,22,23,24,25}. Other regions of the brain, including the cortex, cerebellum, hippocampus, and amygdala, show minimal GLP1R expression in bulk RNA-sequencing.

Single-cell transcriptomic analysis identifies candidate CNS cell populations with detectable GLP1R transcripts

Single-cell RNA-seq analysis of the CNS using the CELLxGENE census obtained from the Chan Zuckerberg Initiative²¹ identified a restricted set of GLP1R-enriched cell types when ranked by GLP1R engagement potential (GEP), defined as mean expression multiplied by the fraction of expressing cells (Fig. 6C). The highest-ranking populations were vascular-associated smooth muscle cells in spinal cord (GEP = 0.0351), contractile cells in spinal cord (0.0213), and GABAergic neurons in spinal cord (0.0183), followed by secretory cells in spinal cord (0.0125), cerebellar granule cell precursors in brain (0.0121), noradrenergic cells in brain (0.0121), perivascular cells in spinal cord (0.0089), Bergmann glial cells (0.0076), cerebellar granule cells (0.0067), cerebellar glutamatergic neurons (0.0066), hippocampal astrocytes (0.0062), and brain vascular cells (0.0055). Together, these analyses identify low-level GLP1R transcript detection across a limited number of CNS cell populations. These observations are descriptive and should not be interpreted as evidence of receptor protein expression, functional signaling, or semaglutide engagement within these cell populations.

This study demonstrated that in the real-world clinical setting, the maximum attained semaglutide dose during the first two years after the first prescription were more closely associated with the subsequent incidence of neuropsychiatric outcomes than achieved weight loss. Specifically, a higher attained dose was associated with lower subsequent incidence of conditions related to mood/affect, anxiety, substance-use, neuromuscular function, and sleep disturbances. In contrast, most of these disease categories did not show significant separation across weight loss strata. This dissociation suggests that the observed neurological associations with semaglutide may not be fully explained by body weight reduction alone and supports the need for future prospective studies, including randomized trials and long-term observational studies during treatment and after treatment cessation.

Given the retrospective nature of this study, it is important to recognize that the observed dose associations reflect more than pure pharmacologic exposure. Maximum prescribed dose is not an independent pharmacological variable but also encompasses multiple factors, including but not limited to tolerability, adherence, access to care, and insurance status. The cohort of patients who did not achieve a higher dose of semaglutide is likely enriched for patients who did not actually begin taking semaglutide, did not tolerate the medication, or discontinued the medication before or during the recommended dose escalation phase. On the other hand, it is likely that the high-dose cohort is generally healthier, more engaged with the healthcare system, and demonstrates higher treatment adherence. Thus, healthy-adherer bias is a possible explanation for the observed associations between higher dose and lower rates of neuropsychiatric outcomes. While weight loss magnitude during the pre-landmark period would also be expected to reflect several of these same treatment adherence-related factors, it is possible that the null findings in the weight loss-stratified analyses simply indicate that the dose axis absorbs a larger component of the healthy-adherer signal, which is not captured by weight loss. In light of this, the dose associations reported in this study should not be interpreted as a dose-response signal in the pharmacological sense, and further randomized prospective studies evaluating neuropsychiatric outcomes in dose-stratified cohorts will be needed to tease apart these axes.

The findings presented in this study should not be interpreted as a benefit–risk assessment or as support for off-label semaglutide prescribing. The present landmark cohort was constructed among patients with preexisting neuropsychiatric disease to evaluate subsequent diagnosis incidence, whereas valid adverse-event analyses require event-specific baseline-negative populations, washout and latency windows, and exposure definitions that distinguish actual initiation, escalation, interruption, and discontinuation. Applying the present post-landmark framework to gastrointestinal, hepatobiliary, pancreatic, metabolic, or acute psychiatric events would be particularly problematic because many such events occur during treatment initiation or dose escalation, within the two-year pre-landmark period, and may themselves prevent attainment of a higher dose. Maximum attained dose, therefore, partly reflects tolerability and persistence, potentially enriching the high-dose group for patients who tolerated and continued treatment. Prior studies using the same federated EHR and AI-curation infrastructure have evaluated 16 label-identified gastrointestinal, systemic, and metabolic adverse events across propensity-matched tirzepatide and semaglutide cohorts and have shown that gastrointestinal adverse effects were among the leading documented reasons for treatment interruption^22,23 {refs}. A prespecified safety analysis in adults with type 1 diabetes additionally identified increases after treatment initiation in nausea or vomiting, hypoglycemia, constipation, diarrhea, decreased appetite, and acute kidney injury, while diabetic ketoacidosis, severe hypoglycemia, pancreatitis, gallbladder disease, gastroparesis, and retinopathy progression did not increase significantly after multiplicity correction; diabetic ketoacidosis was increased among patients with greater than 30% reduction in insulin total daily dose²⁴. These dedicated event-specific studies complement, but do not substitute for, the present hypothesis-generating analysis. Accordingly, the lower incidence of broad neuropsychiatric diagnosis families reported here does not exclude acute psychiatric events and should not be used to infer the absence of treatment-emergent adverse events.

Importantly, this study also highlighted heterogeneity in neuropsychiatric outcomes during and after semaglutide exposure. For example, cognitive symptoms and speech/language symptoms aligned more strongly with weight loss strata than attained dose, and cognitive-symptom event rates increased in the higher achieved weight loss groups. These findings should not be interpreted as evidence that semaglutide worsens cognition, and indeed, prior studies have demonstrated a positive association between weight loss and improved cognitive function²⁵. Rather, like the other results presented here, this trend likely reflects a mixture of biology and bias, including frailty enrichment, illness-related catabolism, differential surveillance, reverse causation, heterogeneous symptom coding, and the clinical reality that the first recorded diagnosis often lags far behind biological onset. Weight loss magnitude is challenging to consider as an exposure variable in retrospective analyses, as weight loss can be driven by both intentional (e.g., pharmacotherapy, lifestyle adjustments) and non-intentional (e.g., known or undiagnosed illnesses, declining reserve, and malnutrition) factors. Future prospective studies will help to disentangle these confounders and more definitively characterize the impacts of semaglutide exposure and therapy-induced weight loss on neuropsychiatric phenotypes.

Existing literature evaluating the association between incretin therapies and neuropsychiatric outcomes has yielded mixed results. Recent clinical trials studying the impact of semaglutide on clinical progression in patients with early Alzheimer’s disease showed no benefit compared to placebo²⁶. In line with this, our analysis did not find any significant difference between the high-dose and low-dose cohorts in the incident diagnosis rates of dementia or residual CNS and neurologic disorders during the post-landmark period. In a retrospective study of nearly 96,000 patients with depression and/or anxiety, semaglutide use was associated with significantly lower risk of worsening depression, worsening anxiety, and worsening substance-use²⁷. Conversely, in a propensity-matched retrospective analysis of over 300,000 obese patients without baseline neuropsychiatric conditions (bipolar disorder, schizophrenia, depression, anxiety, or suicidal ideation), use of GLP-1RA therapy was associated with significantly increased incidence of major depression, anxiety, and suicidal behavior²⁸. Independent of the GLP-1 axis, multiple prior studies have found that intentional weight loss is associated with improvements in some neuropsychiatric outcomes, including depression and cognitive function^25,29. Mechanistically, there is strong evidence that incretin therapies function at least in part via CNS signaling pathways, with preclinical data demonstrating that GLP-1RAs activate hypothalamic neurons and establishing that CNS GLP1R expression is necessary to mediate their appetite suppression, weight loss, and counteracting of “molecular age^30,31,32,33 Whether broader CNS pathways that impact mood, anxiety, personality, and cognition are engaged by these medications warrants further exploration.

The tissue-resolved GLP1R transcriptomic analyses provide descriptive molecular context for the observed clinical associations and identify candidate tissues and cell populations for future investigation. Bulk transcriptomic data identified multiple potential CNS sites of GLP1R expression, including the hypothalamus, caudate, putamen, and nucleus accumbens. These findings are compatible with human neuroimaging studies and recent hypothalamic and hindbrain expression atlases that localize GLP1R to neuronal populations involved in satiety, aversion, interception, and neuroendocrine output^{33,34,35,36,37}. However, the low transcript abundance and observational nature of the analyses preclude inference regarding the cellular mechanisms underlying the clinical associations. The single-cell RNA-seq analyses identified a limited number of CNS cell populations with detectable GLP1R transcripts. The engagement-potential metric was used solely as a descriptive ranking measure and should not be interpreted as evidence of receptor function or pharmacological engagement. While these observations identify candidate tissues and cell populations for future investigation, they do not establish protein expression, receptor signaling, or the biological mechanisms responsible for the clinical associations observed in this study.

This study has several limitations. First, it is a retrospective observational study and is thereby vulnerable to multiple sources of bias, including healthy-adherer and other forms of selection bias, ascertainment bias, and confounding by unmeasured variables. Second, given the large cohorts analyzed in the study, there is adequate statistical power to detect very small differences between the analyzed cohorts. In this setting, there may be statistically significant results that arise for a number of reasons, including true treatment effects and residual confounding despite attempted balancing via propensity-score matching. Thus, it is important to consider not only the level of statistical significance but also the magnitude of difference (effect size) and potential confounding factors that could influence each analysis. Third, disease incidence was defined by the first instance of a diagnostic code documented in the EHR, which inherently reflects diagnostic capture rather than biological onset and is prone to variability based on diagnostic coding practices. Fourth, there are challenges associated with retrospective analysis of both medication dosing (e.g., variable treatment adherence, which is not well captured in the EHR) and weight loss (e.g., selection bias and classification of intentional versus unintentional weight loss). For example, by definition, the low-dose cohort includes patients who never received a prescription for a recommended maintenance dose of semaglutide. It is likely that some of these patients never actually initiated semaglutide or took very few doses before their dose escalation was discontinued. Without access to prescription claims data or detailed narratives of medication adherence, it is difficult to establish the degree of longitudinal medication exposure in this group, which likely varies widely between patients in the group. Fifth, this study did not include an assessment of treatment-related adverse events experienced after semaglutide initiation. The adverse event profile of semaglutide has been characterized in randomized trials and real-world studies, and it is known that adverse events can significantly impact patient adherence and maintenance dose attainment^{38,39,40,41,42}. Further characterization in these dose-stratified patient populations with preexisting neuropsychiatric disease could be informative. Sixth, the broad CNS and symptom-level groupings applied here are analytically useful but clinically heterogeneous. More fine-grained analyses are warranted. Additionally, the analysis indiscriminately considered patients who were prescribed any formulation of semaglutide. These different formulations have shown varying levels of cardiometabolic improvements, with important differences relating to their dose escalation regimens, approval timelines, typical clinical use cases, and underlying characteristics of the patient populations receiving these prescriptions. Regarding the molecular analyses, GLP1R transcript abundance was generally low across CNS tissues evaluated in this study. Consequently, transcript detection may be influenced by sequencing depth, tissue composition, transcript dropout, and cell-type annotation. Furthermore, RNA abundance does not necessarily correlate with receptor protein expression, membrane localization, ligand accessibility, or downstream signaling. Accordingly, these transcriptomic analyses should be interpreted as descriptive and hypothesis-generating rather than mechanistic⁴³.

In summary, this retrospective study identified hypothesis-generating associations between higher attained semaglutide dose and lower subsequent incidence of several broad neuropsychiatric diagnosis families, largely independent of maximum achieved weight loss. These associations may reflect pharmacology, tolerability, treatment persistence, access, healthy-adherer effects, or residual confounding and require rigorous prospective validation.

Data sources and study design

This study was a retrospective observational analysis using de-identified longitudinal electronic health record data from the nference Federated EHR Network²⁰. All analyses were performed on de-identified data. The study integrated semaglutide-associated neurological and neuropsychiatric outcome analyses from the federated EHR platform with GTEx-derived GLP1R mRNA expression data across tissues and sample-level metadata.

The clinical outcomes component comprised summary analyses spanning 24 neurological or neuropsychiatric outcomes, including counts, event-rate ratios, dose-group P values, and weight loss-strata P values. The transcriptomic component comprised GTEx-based GLP1R sample-level gene expression measurements, in transcripts per million (TPM), together with tissue summaries and metadata-stratified association analyses¹².

Study cohorts and exposure definitions

The primary analytic cohort comprised 63,215 semaglutide-treated patients with at least one baseline neurological or neuropsychiatric condition before treatment initiation. Medication exposure was ascertained from prescription or medication event records using the recorded event timestamp. The index date was defined as the date of the first qualifying prescription for the cohort-defining medication or medication class. Follow-up data were available through December 31, 2025. To create incident-user comparator cohorts with minimal treatment contamination, separate mutually exclusive cohorts were defined for metformin, DPP-4 inhibitors, and SGLT2 inhibitors using the same index-date framework and exclusion logic. Other GLP-1RA medications were treated as exclusion exposures when defining the semaglutide cohort.

Demographic and clinical covariates

Demographic variables included age and sex. Age at index was obtained from the index medication record. Baseline body mass index (BMI) was defined as the BMI measurement closest to the index date, recorded from 365 days before through 1 day after the index date. Baseline weight for weight loss analyses was defined separately as the weight measurement closest to the index date, recorded from 90 days before through 14 days after the index date. Baseline type 2 diabetes was defined by the presence of at least three encounters carrying a type 2 diabetes International Classification of Diseases (ICD) code on three distinct dates before the index date. Baseline neurological and neuropsychiatric disease burden was defined using prespecified ICD-9 and ICD-10 code prefixes across 31 clinically relevant domains, including mood/affective disorders, anxiety/trauma/OCD/somatoform disorders, substance-related neuropsychiatric disorders, cognitive symptoms, neurocognitive disorders, speech/language symptoms, neuromuscular disease, cerebrovascular disease, seizure, and paroxysmal neurologic disorders, and hereditary or degenerative CNS syndromes. The union of these domains defined an overall baseline neurological-burden cohort for the semaglutide landmark analyses, whereas individual baseline neurological subgroups were examined separately in comparative analyses. Complete diagnosis code definitions are provided in Table S1.

Propensity-matched comparative neuropsychiatric analyses

For comparator analyses, three mutually exclusive incident-user cohorts were constructed for metformin, DPP-4 inhibitors, and SGLT2 inhibitors using the same index date and exclusion logic described above. Within each baseline neurological or neuropsychiatric subgroup, we compared the rates of each individual incident neuropsychiatric outcome between the semaglutide cohort and the matched cohorts of patients taking metformin, DPP-4 inhibitors, or SGLT2 inhibitors. Propensity scores were estimated with logistic regression using age at index, sex, baseline BMI, and baseline type 2 diabetes status. Patients were matched 1:1 without replacement using nearest-neighbor matching and a caliper of 0.2 on the propensity-score scale.

Follow-up for comparative outcome analyses began on the treatment initiation date. The outcomes were the same as the baseline neurologic and neuropsychiatric diseases listed above. As negative control analyses, we also assessed the incidence of eczema, kidney stones, hernia, rotator cuff disorders, and tinea corporis, which are not expected to be impacted by semaglutide or the other comparator medications. Patients were followed until the outcome of interest, death, or the last observed clinical record. For incident neurological and neuropsychiatric outcomes, patients with a documented diagnosis of the given condition at baseline were excluded from that specific analysis. We summarized matched cohort size, event counts, person-time, and cumulative 2-year event estimates using Kaplan–Meier methods and log-rank P values.

Neuropsychiatric outcomes and landmark analyses

Two-year landmark analyses were conducted within the semaglutide cohort to assess associations of attained dose intensity and achieved weight loss with subsequent neurological and neuropsychiatric outcomes.

For dose-response analyses, each patient’s maximum achieved dose within two years after treatment initiation (defined as the “pre-landmark period”) was paired with that patient’s maximum achieved change in total body weight during the same interval. Dose categories corresponded to labeled dosing increments for injectable semaglutide formulations: 0.25, 0.5, 1.0, 1.7, 2.0, and 2.4 mg or greater. Observed dose values were used when directly available in the medication record, and missing dose values were extracted from medication descriptions using a large language model-assisted dose-mapping workflow⁴⁴.

For post-landmark outcome analyses, the landmark date was defined as the two-year timepoint after treatment initiation for both dose- and weight loss-based analyses. For weight loss-based analyses, maximum percent body weight reduction was calculated from the lowest observed weight recorded from day 15 after index through the landmark date, and patients were categorized into prespecified weight loss strata of <5, 5–10, 10–15, 15–20, 20–25, and ≥25%. For dose-based analyses, the maximum semaglutide dose reached between the index and the landmark date was classified as low dose (0.25–1.0 mg) or high dose (≥1.7 mg). For each incident outcome analysis, patients with evidence of that same condition before follow-up were excluded. Outcomes were the same as defined above for the propensity-matched comparative analyses, including the negative control outcomes. Post-landmark follow-up extended for 24 months and continued until the outcome of interest, death, or the last observed clinical record, whichever occurred first. To compare dose-associated and weight loss-associated patterns across outcomes, the corresponding dose- and weight-stratified P values were jointly visualized in a two-axis bubble plot.

Assessment of GLP1R expression in bulk and single-cell RNA-sequencing datasets

The GTEx-derived GLP1R atlas comprised 18,576 bulk RNA-sequencing samples, with sample-level expression measurements and demographic metadata across 64 tissues. Tissue-level ranking was based on mean TPM values from the tissue summary table. Single-cell RNA-sequencing data were obtained from the CELLxGENE census from the Chan Zuckerberg Initiative²¹. Summary data for all CNS cell types, including the percentage of cells with transcripts detected and the mean expression level in expressing cells were downloaded from the portal. A composite metric termed GLP1R engagement potential (GEP) was defined as the mean expression multiplied by the fraction of expressing cells and was used to rank the list of candidate expressing cell types.

Statistical analysis

Continuous variables were summarized as means with standard deviations or medians with interquartile ranges, and categorical variables were summarized as counts and percentages. In the comparative neuropsychiatric analyses, between-cohort outcome differences were summarized using the Kaplan–Meier method, with significance assessed by log-rank testing; comparative heatmaps displayed absolute differences in approximate two-year Kaplan–Meier event risk. In the semaglutide weight loss landmark analyses, cumulative event distributions and corresponding event-rate comparisons across weight loss strata were assessed using a global log-rank test. In the dose landmark analyses, cumulative event distributions were compared between low-dose and high-dose groups using log-rank tests, whereas post-landmark event-rate comparisons were summarized as relative risks with 95% confidence intervals and Wald-type chi-square P values derived from log risk-ratio estimates. Bubble-plot visualizations displayed dose- and weight loss-associated evidence as -log₁₀(P) values derived from these corresponding dose- and weight-stratified analyses.

For the GTEx analyses, tissue-level and sample-level summary statistics were analyzed descriptively and visualized directly from the available TPM, P value, and q value outputs. For plotting purposes, P values reported as zero in summary exports were truncated to 1 × 10⁻⁶ to enable log-scale visualization. Because the clinical outputs were analyzed in aggregated form and the molecular outputs reflected summary-level bulk-tissue statistics, mechanistic interpretations were treated as hypothesis-generating rather than causal, and no attempt was made to infer mediation or patient-level counterfactual effects.

Data source

This study analyzed de-identified EHR data from academic medical centers in the United States via the nference Clinical Analytics Platform. Prior to analysis, all data underwent expert determination de-identification satisfying HIPAA Privacy Rule requirements (45 CFR §164.514(b)(1)), employing a multi-layered transformation approach for both structured data (cryptographic hashing of identifiers, date-shifting, geographic truncation) and unstructured clinical text (ensemble deep learning and rule-based methods with >99% recall for personally identifiable information detection)⁴⁵. nference established secure data environments within each participating center, housing these de-identified patient data governed by expert determination. These de-identified data environments were specifically designed to enable data access and analysis without requiring Institutional Review Board oversight, approval, or exemption confirmation. Accordingly, informed consent and IRB review were not required for this study.

1. Wilding, J. P. H. et al. Once-weekly semaglutide in adults with overweight or obesity. N. Engl. J. Med. 384, 989–1002 (2021).

   Article  CAS  PubMed  Google Scholar 

2. Lincoff, A. M. et al. Semaglutide and cardiovascular outcomes in obesity without diabetes. N. Engl. J. Med. 389, 2221–2232 (2023).

   Article  CAS  PubMed  Google Scholar 

3. Perkovic, V. et al. Effects of semaglutide on chronic kidney disease in patients with type 2 diabetes. N. Engl. J. Med. 391, 109–121 (2024).

   Article  CAS  PubMed  Google Scholar 

4. Kosiborod, M. N. et al. Semaglutide in patients with obesity-related heart failure and type 2 diabetes. N. Engl. J. Med. 390, 1394–1407 (2024).

   Article  CAS  PubMed  Google Scholar 

5. Jastreboff, A. M. et al. Tirzepatide once weekly for the treatment of obesity. N. Engl. J. Med. 387, 205–216 (2022).

   Article  CAS  PubMed  Google Scholar 

6. Aronne, L. J. et al. Tirzepatide as compared with semaglutide for the treatment of obesity. N. Engl. J. Med. 393, 26–36 (2025).

   Article  CAS  PubMed  Google Scholar 

7. Malhotra, A. et al. Tirzepatide for the treatment of obstructive sleep apnea and obesity. N. Engl. J. Med. 391, 1193–1205 (2024).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

8. Packer, M. et al. Tirzepatide for heart failure with preserved ejection fraction and obesity. N. Engl. J. Med. 392, 427–437 (2025).

   Article  CAS  PubMed  Google Scholar 

9. Loomba, R. et al. Tirzepatide for metabolic dysfunction-associated steatohepatitis with liver fibrosis. N. Engl. J. Med. 391, 299–310 (2024).

   Article  CAS  PubMed  Google Scholar 

10. van Bloemendaal, L. et al. GLP-1 receptor activation modulates appetite- and reward-related brain areas in humans. Diabetes 63, 4186–4196 (2014).

    Article  PubMed  Google Scholar 

11. van Bloemendaal, L. et al. Brain reward-system activation in response to anticipation and consumption of palatable food is altered by glucagon-like peptide-1 receptor activation in humans. Diab. Obes. Metab. 17, 878–886 (2015).

    Article  Google Scholar 

12. GTEx Consortium. The GTEx Consortium atlas of genetic regulatory effects across human tissues. Science 369, 1318–1330 (2020).

13. Tadross, J. A. et al. A comprehensive spatio-cellular map of the human hypothalamus. Nature 639, 708–716 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

14. Ludwig, M. Q., Todorov, P. V., Egerod, K. L., Olson, D. P. & Pers, T. H. Single-cell mapping of GLP-1 and GIP receptor expression in the dorsal vagal complex. Diabetes 70, 1945–1955 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

15. Huang, K.-P. et al. Dissociable hindbrain GLP1R circuits for satiety and aversion. Nature 632, 585–593 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

16. Zhu, Z. et al. Hedonic eating is controlled by dopamine neurons that oppose GLP-1R satiety. Science 387, eadt0773 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

17. Li, Z. et al. Systemic GLP-1R agonist treatment reverses mouse glial and neurovascular cell transcriptomic aging signatures in a genome-wide manner. Commun. Biol. 4, 656 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

18. Venkatakrishnan, A. J., Murugadoss, K. & Soundararajan, V. Weight loss dynamics and health burden changes with tirzepatide versus semaglutide. PNAS Nexus 5, pgag171 (2026).

19. Murugadoss, K., Varma, G., Venkatakrishnan, A. J., Gibson, M. C. & Soundararajan, V. Weight trajectories after last tirzepatide or semaglutide prescription across a federated health network. Biol. methods protoc. 11, bpag020 (2026).

20. Venkatakrishnan, A. J. et al. Clinical nSights: a software platform to accelerate real world oncology analyses. J. Clin. Oncol. https://doi.org/10.1200/JCO.2024.42.16_suppl.e23316 (2024).

21. Cell Science Program, C. Z. I. et al. CZ CELLxGENE Discover: a single-cell data platform for scalable exploration, analysis and modeling of aggregated data. Nucleic Acids Res. 53, D886–D900 (2025).

    Article  Google Scholar 

22. Venkatakrishnan, A. J., Murugadoss, K. & Soundararajan, V. Weight-loss dynamics with tirzepatide versus semaglutide. PNAS Nexus 5, gag171 (2026).

    Article  Google Scholar 

23. Varma, G. et al. Auditable large language model curation of clinical notes refines glucagon-like peptide-1 initiation, persistence ascertainment, and compounding use beyond prescription records. Biol. Methods Protoc. 11, bpag035 (2026).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

24. Matson, R., Murugadoss, K., Venkatakrishnan A. J. & Soundararajan V. Semaglutide and Tirzepatide in Type 1 Diabetes: Real-World Insulin Deintensification, Cardiovascular Outcomes and Safety Assessment. https://www.preprints.org/manuscript/202606.0638 (2026).

25. Veronese, N. et al. Weight loss is associated with improvements in cognitive function among overweight and obese people: a systematic review and meta-analysis. Neurosci. Biobehav. Rev. 72, 87–94 (2017).

    Article  PubMed  Google Scholar 

26. Cummings, J. L. et al. Efficacy and safety of oral semaglutide 14 mg (flexible dose) in early-stage symptomatic Alzheimer’s disease (evoke and evoke+): two phase 3, randomised, placebo-controlled trials. Lancet https://doi.org/10.1016/S0140-6736(26)00459-9 (2006).

27. Taipale, H. et al. Association between GLP-1 receptor agonist use and worsening mental illness in people with depression and anxiety in Sweden: a national cohort study. Lancet Psychiatry 13, 327–335 (2026).

    Article  PubMed  Google Scholar 

28. Kornelius, E., Huang, J.-Y., Lo, S.-C., Huang, C.-N. & Yang, Y.-S. The risk of depression, anxiety, and suicidal behavior in patients with obesity on glucagon like peptide-1 receptor agonist therapy. Sci. Rep. 14, 24433 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

29. Fabricatore, A. N. et al. Intentional weight loss and changes in symptoms of depression: a systematic review and meta-analysis. Int. J. Obes. 35, 1363–1376 (2011).

    Article  CAS  Google Scholar 

30. Chen, Z. et al. GLP-1R-positive neurons in the lateral septum mediate the anorectic and weight-lowering effects of liraglutide in mice. J. Clin. Invest. 134, e178239 (2024).

31. Huang, J. et al. Body-wide multi-omic counteraction of aging with GLP-1R agonism. Cell Metab. 37, 2362–2380.e8 (2025).

    Article  CAS  PubMed  Google Scholar 

32. Polex-Wolf, J. et al. coexpressing neurons modulate the suppression of food intake and body weight by a GLP-1/leptin dual agonist. Sci. Transl. Med. 16, eadk4908 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

33. Mazzaferro, S. et al. GLP-1R agonists activate human hypothalamic neurons. Preprint at bioRxiv https://doi.org/10.1101/2024.04.02.587825 (2026).

34. Burmeister, M. A. et al. The hypothalamic glucagon-like peptide 1 receptor is sufficient but not necessary for the regulation of energy balance and glucose homeostasis in mice. Diabetes 66, 372–384 (2017).

    Article  CAS  PubMed  Google Scholar 

35. Kim, K. S. et al. GLP-1 increases preingestive satiation via hypothalamic circuits in mice and humans. Science 385, 438–446 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

36. Psilopanagioti, A. et al. Glucagon-like peptide-1 receptor in the human hypothalamus is associated with body mass index and colocalizes with the anorexigenic neuropeptide nucleobindin-2/Nesfatin-1. Int. J. Mol. Sci. 23, 14899 (2022).

37. López-Ferreras, L. et al. Lateral hypothalamic GLP-1 receptors are critical for the control of food reinforcement, ingestive behavior and body weight. Mol. Psychiatry 23, 1157–1168 (2017).

    Article  PubMed  PubMed Central  Google Scholar 

38. Shu, Y. et al. Gastrointestinal adverse events associated with semaglutide: a pharmacovigilance study based on FDA adverse event reporting system. Front. Public Health 10, 996179 (2022).

    Article  PubMed  PubMed Central  Google Scholar 

39. Gasoyan, H. et al. Reasons for discontinuation of obesity pharmacotherapy with semaglutide or tirzepatide in clinical practice. Obesity 33, 2296–2303 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

40. Brufani, C. & Morviducci, L. Semaglutide in a real-world outpatient setting: discontinuation patterns and weight maintenance. Obesity Endocrinol 1, wjaf011 (2025).

41. Rodriguez, P. J. et al. Discontinuation and reinitiation of dual-labeled GLP-1 receptor agonists among US adults with overweight or obesity. JAMA Netw. Open 8, e2457349 (2025).

    Article  PubMed  PubMed Central  Google Scholar 

42. Nauck, M. A. et al. A phase 2, randomized, dose-finding study of the novel once-weekly human GLP-1 analog, semaglutide, compared with placebo and open-label liraglutide in patients with type 2 diabetes. Diab. Care 39, 231–241 (2016).

    Article  CAS  Google Scholar 

43. Edfors, F. et al. Gene-specific correlation of RNA and protein levels in human cells and tissues. Mol. Syst. Biol. 12, 883 (2016).

    Article  PubMed  PubMed Central  Google Scholar 

44. Website. https://doi.org/10.64898/2026.04.02.26350077.

45. Murugadoss, K. et al. Building a best-in-class automated de-identification tool for electronic health records through ensemble learning. Patterns 2, 100255 (2021).

    Article  PubMed  PubMed Central  Google Scholar 

46. OpenAI et al. GPT-4o system card. (2024).

Download references

Authors and Affiliations

1. nference, 1 Main Street, East Arcade Building, Cambridge, Massachusetts, MA, USA

   Karthik Murugadoss, A. J. Venkatakrishnan & Venky Soundararajan

2. Metabolism Agentic Intelligence Atlas (MAIA), Cambridge, Massachusetts, MA, USA

   Karthik Murugadoss, A. J. Venkatakrishnan & Venky Soundararajan

Contributions

V.S. conceived and designed the study. K.M and A.J.V. contributed software tools. V.S., A.J.V, and K.M. drafted the manuscript. All authors interpreted the data, revised the manuscript critically for important intellectual content, and approved the final version.

Corresponding author

Correspondence to Venky Soundararajan.

Competing interests

The authors are employees of nference, inc., which conducts research collaborations with various biopharmaceutical companies whose therapeutic products are included in this study. None of these companies, nor any other nference collaborator, funded, supported, or had any role in the independent study design, data acquisition, analysis, interpretation, manuscript preparation, or the decision to submit this work for publication. All analyses were conducted by the authors using de-identified electronic health record data. The authors declare no additional competing interests.

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions