Figure 1.

Associations of the plasma metabolome with the steatosis grade

(A, D, G, and J) Boxplots of the normalized permutated variable importance measure (VIM) for the metabolites associated with the liver steatosis grade after controlling for age, BMI, gender, and country in the discovery cohorts (n = 117, biopsy; n = 111, echography), validation cohort 1 (n = 263, echography), and validation cohort 2 (n = 271, echography), respectively. Metabolites were identified using the random forest-based ML variable selection algorithm Boruta using 2,000 trees, 500 iterations, and p_Bonferroni < 0.005.

(B, E, H, and K) SHAP summary plot of the metabolites associated with the liver steatosis grade selected by the Boruta algorithm in the discovery (biopsy and echography), validation 1 (echography), and validation 2 (echography) cohorts, respectively. Each dot represents an individual sample. The x axis represents the SHAP value: the impact of a specific metabolite on the liver steatosis grade prediction of a specific individual. The overall importance for final prediction (average absolute SHAP values) is shown in bold. Colors represent the values of the metabolite normalized concentrations, ranging from yellow (low concentrations of the specific metabolite) to purple (high concentrations of the specific metabolite).

(C, F, I, and L) Violin plots showing the ranked residuals (after adjusting for age, BMI, gender, and country) of plasma histidine levels according to the degree of steatosis assessed by liver biopsy in the discovery cohort (n = 117), liver echography in the discovery cohort (n = 111), liver echography in validation cohort 1 (n = 263), and liver echography in validation cohort 2 (n = 271), respectively. Statistical significance was assessed using both Kruskal-Wallis and Mann-Kendall trend tests, and between-group comparisons were assessed using the Wilcoxon test. #p < 0.1, ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.