*3.3. SO-CovSel Analysis*

Serum levels of 64 biomolecules, including cytokines, chemokines, growth factors, amino acids, and derivatives, and the abundance of 77 gu<sup>t</sup> microbial taxa, were assayed through multiple analytical platforms. Concentrations of serum mediators are reported in Table S2. Several SO-CovSel models were built using the multimatrix dataset. Among all of the tested models, the one that allowed the best classification of PF&S and nonPF&S participants with the smallest number of variables was selected. This latter was built using only seven analytes (Table 4).


**Table 4.** Levels of relevant analytes as resulted from SO-CovSel analysis.

Data are shown as median and interquartile range.

The rate of correct classification was 91.7% for PF&S participants and 87.5% for the nonPF&S group (90.0% in the whole study population), with an average area under the receiver operating characteristic (AUROC) curve very close to 1. When compared with their distributions under the null hypothesis, all of the classification figures of merit were statistically significant (*p* < 0.0001). Among the discriminant analytes selected by the SO-CovSel model, participants with PF&S showed lower levels of MIP-1β. As for the metabolic signature, the level of the aspartic acid was higher in PF&S participants, while that of threonine was higher in the nonPF&S group. Among the gu<sup>t</sup> microbes contributing to the model, *Oscillospira* and *Ruminococcus* were more abundant in PF&S participants, while Barnesiellaceae *and* Christensenellaceae were higher in the nonPF&S group (Table 4).
