Microbiome-Based Body Mass Index Prediction : A Scalable Machine Learning Benchmark for Large-Scale Metagenomic Data
avoin
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
Lataukset18
Pysyvä osoite
Verkkojulkaisu
DOI
Tiivistelmä
The human gut microbiome is closely linked to host metabolism. Although many studies have reported associations between specific microbial taxa and obesity, the extent to which gut microbiome composition can predict body mass index (BMI) at population scale remains uncertain. This thesis investigates microbiome-based BMI prediction using 17,903 human gut metagenomes from the Metalog consortium and evaluates whether predictive performance improves through non-linear machine-learning models, large sample sizes, and substantial feature reduction.
The analysis was implemented as a reproducible Nextflow pipeline designed for portability and feasibility on consumer-grade hardware; an independent Snakemake implementation was additionally developed as a reproducibility check, although the two pipelines were not run with identical procedures and so are not compared one-to-one. Linear models (Elastic Net and linear SVM) were compared against non-linear tree-based methods. Random Forest consistently achieved the strongest performance, reaching an RMSE of 5.06 kg/m2 (R2 = 0.387), whereas the linear baselines remained substantially weaker. Saturation analyses indicated a region of diminishing returns between approximately 12,000 and 14,000 samples, with only a small further gain at the largest tested size (N=16,000). A descriptive analysis showed that the extreme BMI values reflect participant age and clinical cohort (infants at the low end; bariatric, elderly, and metabolic-disease cohorts at the high end) rather than data-entry error, and that top-ranked species show only weak, distributed associations with BMI.
To address the high dimensionality and sparsity of the data, a 1% prevalence filter and a variance-based top-500 filter were evaluated; these substantially lowered runtime and memory requirements while preserving most of the predictive signal. Feature-ranking analyses identified a mixture of named species and uncharacterised species-level genome bins among the highest-ranked predictors, including Bifidobacterium longum, Veillonella species, Ruminococcus gnavus, and several uncharacterised SGBs (e.g., GGB9512_SGB14909) — consistent with MetaPhlAn 4's improved profiling of previously uncharacterised taxa. Overall, the thesis establishes a scalable and reproducible baseline for microbiome-based BMI prediction while clarifying methodological trade-offs in evaluation design, preprocessing, and feature selection.