23andMe Unveils Methodology for Investigating GLP1 Receptor Agonist Efficacy and Side Effects

A recent study by 23andMe has detailed its comprehensive methodology for evaluating the efficacy and adverse reactions associated with GLP1 receptor agonist medications. The research combines self-reported data from 23andMe customers with genetic and Electronic Health Record (EHR) information.

The primary objective of this study is to pinpoint both genetic and non-genetic factors that may influence how effective GLP1 medications are and the frequency of side effects.

Study Design and Participant Recruitment

Participants for this extensive study were drawn from the 23andMe customer base, each providing informed consent under a protocol rigorously approved by the Salus Institutional Review Board.

The 23andMe GLP1 survey, launched in August 2024, was specifically designed for individuals with a history of using prescription weight-loss medications. This survey gathered in-depth information, including drug brand, specific dosing, total treatment duration, self-reported efficacy (comparing pre-treatment and on-treatment weight), detailed side effects, and the precise reasons for either starting or discontinuing GLP1 treatment.

The analysis focused on a selection of six key drug types:

• Ozempic
• Wegovy
• Compounded semaglutide
• Mounjaro
• Zepbound
• Compounded tirzepatide

Data Collection and Phenotype Definitions

Efficacy

Efficacy was precisely defined as the percentage change in Body Mass Index (BMI) from the pre-treatment baseline to post-treatment or current BMI. For participants who had used multiple GLP1 medications, the drug taken for the longest duration was systematically chosen for analysis.

The percentage BMI change (ΔBMI %) was calculated using the formula:
$${{\rm{\Delta BMI}}}{ % }=100({\mathrm{BMI}}{2}-{\mathrm{BMI}}{1})/{\mathrm{BMI}}{1}$$
Rigorous quality control filters were applied to participant data, specifically for weight, height, BMI, and age. Outlier ΔBMI % values (those exceeding 20% or falling below -45%) were carefully excluded from the analysis. A corresponding Δweight phenotype, representing the change in weight from baseline in kilograms, was also defined, which is mathematically equivalent to ΔBMI % given a constant adult height.

Side Effects

Side effect phenotypes were established as clear case-control classifications for each reported adverse reaction. Participants who self-rated their side effects as moderate or severe were categorized as "cases," while those who reported mild or non-existent side effects served as "controls." Similar to the efficacy analysis, the longest-taken GLP1 medication was prioritized for defining these side effect phenotypes.

Covariates

Additional covariates integrated into the study included:

• Drug type (categorized as semaglutide or tirzepatide)
• The most recent weekly dosage in milligrams
• The total number of days on treatment

Data Comparison and Non-Genetic Factors

Electronic Health Record (EHR) data, voluntarily shared by 23andMe participants through Apple HealthKit, were leveraged for crucial comparison with the self-reported survey data.

To meticulously analyze the influence of non-genetic factors on BMI loss, a linear model was meticulously fitted for ΔBMI %. This comprehensive model incorporated several key variables: age, sex, initial BMI, drug type, dose, days on treatment, and various interaction terms, all while carefully accounting for standard dosing differences between semaglutide and tirzepatide.

Genotyping and Association Testing

DNA was meticulously extracted from saliva samples and genotyped using various advanced Illumina platforms. Participant genotype data were then imputed against a comprehensive reference panel, which included data from the Haplotype Reference Consortium.

A Genome-Wide Association Study (GWAS) of ΔBMI % was conducted in a substantial cohort of 15,237 unrelated participants of European ancestry, following stringent filtering for complete data and relatedness. This GWAS incorporated several important covariates: age, sex, initial BMI, drug type, dose, days on treatment, five genetic principal components, and genotyping platform indicator variables. A similar GWAS procedure was systematically applied for the various side effect phenotypes. Furthermore, drug-specific GWASs were also performed separately for semaglutide-treated and tirzepatide-treated populations. For non-European populations, association testing primarily focused on variants that had been identified in the European GWAS due to smaller available sample sizes.

Replication Efforts

Replication of identified efficacy associations was diligently attempted in two independent cohorts:

• All of Us cohort: This effort utilized EHR data and specific drug codes for semaglutide or tirzepatide from a sample of 3,948 participants with complete data. A larger sample of 4,855 participants was subsequently analyzed after mean-imputation of missing drug dose data.
• UK Biobank cohort: An attempt at replication was also made in this cohort; however, the available data predated the widespread use of semaglutide or tirzepatide, necessitating reliance on earlier GLP1 receptor agonists for comparison.

Genetic and Non-Genetic Risk Modeling

Combined genetic and non-genetic models were developed with the aim of predicting ΔBMI % and the potential risk of treatment-related side effects. Predictors incorporated a wide range of variables including treatment specifics, clinical metrics, demographic information, disease diagnoses (such as type 2 diabetes, hypertension, and non-alcoholic fatty liver disease), genetic variables, and years of education.

A linear multi-variable model was employed for the prediction of ΔBMI %, while multi-variable logistic regression models were utilized for the binary side effect phenotypes. The dataset was judiciously partitioned into a 70% training set and a 30% test set for robust model development and subsequent performance assessment. Efficacy model performance was further evaluated on a separate sample of 642 HealthKit EHR participants who had not completed the GLP1 survey, with unknown variables imputed to simulate a realistic pre-treatment prediction scenario.