Using label propagation classifiers to assess the generalizability of the prospective metabolomic endotypes to other timepoints

We recapitulated the prospective metabolomic endotypes with metabolomic profiling from plasma samples collected during/after COVID infection using a label propagation classifier approach,³¹ This approach applies the classifier by introducing individuals one at a time into the original similarity networks and assigning them to one of the defined metabolomic endotypes and used in previous studies.^20,32 For individuals with multiple plasma samples, a majority voting scheme was applied to determine the final metabolomic endotype membership. In the rare event of a tie, we choose the metabolomic endotype membership based on the COVID severity of the individuals. To examine the associations between recapitulated metabolomic endotypes and clinical variables, such as demographic characteristics, lab tests, comorbidity diseases, COVID-19 Severity, and Influenza Incidence/Recurrence, we conducted a similar analysis as we did for the prospective metabolomic endotypes. This involved performing ANOVA or Chi-square tests to assess whether clinical variables exhibited differentiation among the recapitulated metabolomic endotypes. Additionally, regression analyses with the same adjusted covariates were employed to identify the recapitulated severe endotype that significantly differed from the other endotypes. The endotypes are considered generalizable when consistent ID responses are observed between the prospective and recapitulated metabolomic endotypes.