Development of a metabolic signature of post-weaning diarrhoea in pigs

In the agricultural sector, antibiotics have been used to improve swine growth performance. This application is banned nowadays, due to increased risk of antibiotic resistance. In piglets this results in a higher prevalence of post-weaning diarrhoea, deteriorating both animal health and performance. Our goal was to find a metabolite signature separating piglets with low faecal consistency score (FCS) from piglets with normal faecal consistency and determine which pathways were enriched in this signature. By using direct infusion mass spectrometry on blood spots, we built machine learning (ML) models that aimed to differentiate between low and normal FCS. To test the general predictive capability of these models, we applied a Leave-One-Country-Out (LOCO) strategy for cross validation. Our second approach after LOCO was finding the optimal number of features to include in a feature-reduced model. To determine the order in which features were to be eliminated, we ranked them based on a combination of t-test and fold-change significance scores. Enrichment analysis using mummichog was used to gain insights into the final signature set of m/z values found using this ranking and ML models. Models trained both using all countries and leaving out specific countries from training showed a limited ability to predict FCS category. Furthermore, the LOCO results were mixed, with some countries showing a predictive signal present in the data, but others with predictive capability that was no better than random. Signature analysis using t-test and fold-change results did not result in any KEGG pathways that were enriched in this signature as compared to random. Using these methods, we could predict the FCS category to a limited degree. Although no common signature for low faecal consistency could be determined using this method, given that some countries showed more reliable LOCO results, further analysis into specifically the samples from those countries could be a valuable next step. No metabolite signature enriched in changes to KEGG pathways was found in the data using the combined t-test and fold-change analysis ranking method. We have shared the data with the community through the MetaboLights repository.