Early-life proteomic and microbiome features signal obesity risk across 26 years of follow-up

Figures

• 

  Fig 1. Study design. This cohort study is based on the All Babies in Southeast Sweden (ABIS) general population study (N = 16,683). Biomarkers were assessed for association with future obesity outcomes, including body mass index (BMI) classifications based on International Obesity Task Force (IOTF) standards and clinical diagnoses (ICD-10 codes E66.0, E66.8, and E66.9) through age 26. Biomarkers included metabolomics and targeted proteomics from cord blood, as well as microbiome sequencing from stool samples collected at approximately 1 year of age. Wilcoxon tests and differential expression analysis from binomial distribution modeling (via DESeq2) identified top metabolites, proteins, and taxa associated with obesity outcomes. Environmental and lifestyle exposures from pregnancy to the first year were modeled using machine learning (ML) with a 70/30 train-test split. The final integrated data set included 1,877 individuals with complete biological samples, which were used in the XGBoost ML pipeline. Recursive feature elimination (RFE) was employed for dimension reduction, and L1 regularization with k-fold cross-validation was applied to predict obesity diagnosis up to age 26. More…

• 

  Fig 2. Analytical process. The pipeline for the investigation, demonstrating the metabolomic (Me), proteomic (P), and microbial (Mi) data sets used in each analysis.

• 

  Fig 3. Early-life factors associated with future clinical obesity diagnosis or obesity classification according to IOTF standards (ages 3–19 years). ORs and 95% confidence intervals (CIs) are provided for each reference group and comparison, stratified by IOTF classifications and future obesity diagnosis (up to December 2023). Factors associated with (A) obesity classifications based on BMI at 3, 5, 8, 10, and 19 years of age against IOTF standards, and (B) clinical diagnosis of obesity from the medical record. ORs represent unadjusted associations between exposure and outcome.

• 

  Fig 4. Proteomic differences in cord blood based on childhood BMI classification. Wilcoxon test statistics comparing NPX values of proteins in cord blood, based on BMI classifications across time. Controls are defined as normal-weight by IOTF standards, based on a parent-completed questionnaire, and with no future autoimmune or neurodevelopmental diagnosis from the Swedish National Patient Register. Comparisons include the following: 2–10 years global comparison (218 controls vs. 72 a BMI corresponding to obesity at least once across the timepoints), BMI at 2 years (129 controls vs. 45 overweight/obese), BMI at 5 years (92 controls vs. 51 overweight/obese), and BMI at 8 years (43 controls vs. 19 overweight/obese). (A) Children with BMIs indicative of obesity per IOTF standards at any point between 2 and 3, 5, 8, or 10 years of age were compared to controls using Wilcoxon tests with FDR correction. (B–H) Children classified as overweight or obese per the IOTF definition were compared to controls: (B–D) at age 2 years, (E) at age 5 years, (F–H) at age 8 years. No proteomic findings in panels A or E reached significance after FDR correction. See Table S1 for statistics.

• 

  Fig 5. Birth proteome and metabolome associations with future clinically diagnosed obesity. (A) Wilcoxon test results for cord blood protein levels in relation to future ICD-10 diagnosis of clinical obesity (identified in the Swedish National Patient Register), corrected by FDR correction. (B–K) NPX levels for top hits, stratified by future obesity diagnosis case status. See Table S2 for statistics. (L) Top 15 proteins contributing to case–control discrimination of obesity (obesitas), based on mutual information (MI) scores. (M) Predictive performance of ML and logistic regression models predicting future obesity diagnosis using the top 15 proteins from panel K. (N) Improved AUC-ROC (means and standard deviation) after incorporating four additional proteins—ANGPTL4, FST, HGF, and TNFRSF11B—identified in earlier obesity comparisons. (O–P) Concentrations of cord serum metabolites differed between normal-weight and overweight infants, with density curves representing the concentration distributions.

• 

  Fig 6. Differentially abundant or prevalent bacterial species by future diagnosis of obesity. Differentially abundant or prevalent bacterial species in future clinical diagnosis of obesity. (A and B) Significant differences in bacterial species abundance, after FDR correction. Controls were defined as those without obesity at the 2–3 years and 5 years timepoints, based on their BMI classifications according to the IOTF standards, as well as no future autoimmune or neurodevelopmental diagnosis. A total of 598 controls were compared to 50 future clinical obesity cases. The 16s rRNA sequencing data were assessed for differentially abundant species between obesity and control groups using DESeq2, with a binomial distribution employed for modeling the counts and using a local fit type to adjust for dispersion differences within the data set. This is important for the heterogeneous variance seen in these data sets. Multiple testing was adjusted for using the FDR method. (C) Significant differences in the prevalence of bacterial species. Chi-square tests assessed differences in bacterial prevalence between control and obesity groups, with Fisher’s exact tests used for taxa with low counts. For…

• 

  Fig 7. Integrated machine learning models predicting future clinical diagnosis of obesity. Prediction of clinical obesity diagnosis in 1,877 individuals, based on the XGBoost Classifier with an 80/20 train/test split and RFE integrated within fivefold stratified cross-validation. The primary outcome was clinical diagnosis in the Swedish NPR (22). The integrated predictive models relied exclusively on registry-confirmed clinical obesity diagnoses as the primary outcome, thereby avoiding dependence on self-reported anthropometric measurements. Panels A and B show predictions based on proteins at birth and taxa present in one-year stool samples, while panels C and D additionally incorporate metadata from the birth and 1-year questionnaires. For each fold, the training set was used to select the top 40 features via RFE (from a total of 683 features, including 230 bacterial and 363 proteomic). The model was subsequently trained on these features and evaluated on the held-out test set. L1 regularization (alpha = 1) and class weighting (scale_pos_weight) were applied to handle potential overfitting and class imbalance. Missing values were handled using median imputation fit on the training…