Longitudinal Transcriptomic Remodeling of Adipose Tissue After Bariatric Surgery Revealed by Differential Expression and Explainable Machine Learning

Figures

• 

  Figure 1. Overview of the statistical and machine-learning workflow for gene expression analysis following bariatric surgery.

• 

  Figure 2. Global transcriptomic remodeling following RYGB surgery. (a,b) Principal component analysis (PCA) score plots showing separation between baseline and post-surgery samples for before surgery vs. 2 years post-surgery (a) and before surgery vs. 5 years post-surgery (b). (c,d) Volcano plots showing differentially expressed genes for before surgery vs. 2 years post-surgery (c) and before surgery vs. 5 years post-surgery (d). The x-axis represents log2 fold change, and the y-axis represents −log10(p-value); significantly upregulated and downregulated genes are highlighted. (e,f) PLS-DA loading plots showing the genes contributing most strongly to group separation for before surgery vs. 2 years post-surgery (e) and before surgery vs. 5 years post-surgery (f). (g,h) Combined machine-learning gene-importance rankings based on averaged feature-importance scores from tree-based models for before surgery vs. 2 years post-surgery (g) and before surgery vs. 5 years post-surgery (h).

• 

  Figure 3. The figure displays data for a selection of the most significantly (A) upregulated and (B) downregulated genes. For each gene, boxplots (left) show normalized expression for baseline (Pre, n = 50), 2-year (2Y, n = 49), and 5-year (5Y, n = 38) groups. Boxes denote the interquartile range (IQR) and whiskers 1.5× IQR. Statistical significance was determined by a paired t-test with FDR correction (****, p < 0.0001). Corresponding ROC curves (right) illustrate each gene’s individual classification performance, with the Area Under the Curve (AUC) noted.

• 

  Figure 4. Concordance of KEGG pathway enrichment between analytical methods. The dot plot contrasts the top KEGG pathways derived from the top-ranking genes at 5 years post-surgery. Pathways were identified using two gene-selection strategies: a bioinformatics approach based on differential-expression analysis (“Bio”, blue) and a machine-learning–based approach (“ML”, red). The x-axis represents GeneRatio, dot size indicates pathway gene count, and dot color represents adjusted p-value. The substantial overlap of immune- and inflammation-related pathways demonstrates strong concordance in biological interpretation between differential-expression analysis and machine-learning–based gene prioritization.

• 

  Figure 5. SHAP-based interpretation of the final best-performing classifiers. The baseline-versus-2-year classifier included the seven top-ranked features that produced the best classification performance, whereas the baseline-versus-5-year classifier included the optimal three-feature subset. SHAP values were calculated using these final models and their exact input feature sets. Accordingly, the plots display all features included in each classifier rather than selected top-N features from a 20-gene model. Positive SHAP values shift predictions toward the post-surgery class, whereas negative values shift predictions toward the baseline class.