Marilen Benner

CHAPTER 7 194 DISCUSSION In utero exposure to antibiotics could impact neonatal immunity through two different routes: by disrupting microbial colonization of the child, needed for healthy immune development, or by the antibiotics’ impact on maternal immunity. The latter is associated with pregnancy complications known to alter offspring development (38-41). In the current study, we examined the effect of antibiotic treatment on maternal immunity using a murine gestational model. The chosen microbial intervention strategy was shown to affect offspring immunity (31). In line with this previous investigation of this combination of antibiotics, antibiotic treatment significantly decreased microbial diversity in the ceca of pregnant mice, but none of the mice were observed to have any clinical symptoms or reduced reproductive success. It was however possible to detect a maternal shift in immunological profile by using a machine learning ensemble classification strategy to assess >100 immune parameters of different sampling sides simultaneously (33, 34). Recursive feature elimination reduced the assessed parameters to a selection of 4 immunological features that distinguished control from antibiotic treated mother animals with an accuracy of >90%. Of note, these immune adaptations were observed throughout the maternal body, reaching as far as the placenta. Antibiotics are known to affect intestinal immunity due to microbial manipulation of this immunological compartment (17, 42). Little is known on how this translates to other immune sites, especially regarding gestational tissues. In this exploratory approach, samples of placenta and amniotic fluid were included, and immune features were obtained through multiple methods, i.e. a combination of flow cytometry, mRNA expression, and analysis of soluble factors, to cover a wide range of possible effects. Using machine learning and recursive feature selection allowed for an open assessment of studied parameters. This dimensionality reduction enabled an unbiased focus on the strongest induced changes. Additionally, the applied ensemble strategy offered stable feature selection in this low sample size setting. After mating, of the 40 females that were randomly allocated to the control and treatment cohort, 11 and 8 mice respectively were pregnant and available for a final readout. Small sample size and large variation of the input data could weaken reproducibility when single selection algorithms are used, a limitation that could be overcome by the applied ensemble approach (34, 43). The presented high accuracy of the classification underlines how computational methods may help to reduce the number of animals needed for in vivo studies. While earlier findings showed that antibiotic treatment resulted in a macrophage-dependent increase in inflammatory colonic Th1 responses mice (17), we did not observe any differences in Th lymphocyte populations in the MLN, the gut-draining lymph nodes, of the antibiotic treated animals. Of note, the study by Scott and colleagues investigated immunity in male mice (17) and thus could not take into account the highly specialized immune dynamics of