Multi-omics analysis

Microbiome prediction of longitudinal phenotypes across time space and host

---

Abstract

The microbiome is increasingly viewed as an intermediate layer mediating genetic and environmental effects on complex traits. Integrating microbiome profiles into prediction models may capture latent information and improve phenotype prediction. However, the stability of this contribution across time, seasons, and environments remains unclear.We analyzed 513 dairy animals from 11 farms, each sampled for fecal microbiome (shotgun metagenomics) at three time points: T1 (summer 2024), T2 (winter 2025), and T3 (summer 2025). Genomic, physiological, and environmental data were available for all individuals. Test-day milk yield (kg/day) was the target phenotype; for each time point, we considered the closest test-day records (≈6 per animal within ±4 months).We implemented a hierarchical model in which the microbiota was used as the sole effect, allowing its contribution to be partitioned into microbiome-mediated genetic and environmental components. Its performance was compared with a baseline model including only direct genetic and environmental effects.Three prediction scenarios were tested: Within time point (T1→T1): for each farm, 80% of the information at T1 were used in training and the remaining 20% were used for testing to predict individual phenotypes using their microbiome profiles.Across time (T1→T2 and T1→T3): the model was trained on the full T1 dataset and then used to predict phenotypes at later time points (T2 and T3) using the animals' microbiome profiles from those time points.Across time environments: predictions for two herds excluded from training, under both within- and across-time designs.Prediction accuracy (correlation between predicted and observed phenotypes) was evaluated in 10-day intervals based on the time gap between microbiome and phenotype sampling.Models integrating microbiome data improved prediction accuracy compared with the baseline model. When trained on T1, correlations were 0.60 for T1→T1, 0.43 for T1→T2, and 0.54 for T1→T3, whereas the baseline model achieved ~0.20. Accuracy declined exponentially with increasing time lag between microbiome and phenotype sampling, approaching zero beyond four months. Interestingly, although reduced, microbiome-based predictions remained superior to the baseline even across different environments and seasons.Using the hierarchical model, it was possible to decompose the direct effect of the microbiota on the phenotype and its predictive contribution. Results showed that the loss of predictive power over time was driven by a decay of the microbiome-mediated environmental component, whereas the microbiome-mediated genetic component remained proportionally stable. Additionally, we observed that prediction across herds was still achievable, particularly when focusing on the non-mediated component of the phenotype (together with the genetic component). In summary, we demonstrated that: (i) microbiome data enable phenotype prediction in new environments even several months after sampling, and (ii) they provide insight into how the microbiome influences the phenotype and why its contribution varies across time and space.

Keywords: 2026

How to Cite:

Sartori, C., Oian, A., Rulli, E., Gomez Proto, G., Giagnoni, L., Squartini, A., Treu, L., Campanaro, S., Mantovani, R. & Mancin, E., (2026) “Microbiome prediction of longitudinal phenotypes across time space and host”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2286303. doi: https://doi.org/10.31274/wcgalp.23914

Rights: 1