Estimation & Prediction

Delivering Machine Learning Approaches to predict Methane Emission Production via Microbial Features

---

Abstract

Global climate change is a pressing concern affecting the animal production industry. The recent advent of cheap, high-performance computing, and affordable sequencing has enabled opportunities to use Machine Learning (ML) to predict methane (CH4) emission production in dairy cows. Our aim is to create a set of ML models to examine the relationship between CH4 emission production, host genome and microbiome. To do this, we used 45 samples containing CH4 records, environmental data, and relative microbial abundances from 15 Holstein- Friesian cows to train several Regression and Classification-based ML models with 5,10,15,20 cross-validations (CV). This work serves as the basis for ML algorithms involving the use of cow genotype for prediction, thus paving the way for further research. We used Qiime 2 to prepare the metagenomic data, and R to prepare and process the management data, as well as to combine the metagenomic and management data together. Four ML algorithms were used. These included two regression algorithms: Linear Regression (LM) and Generalised Linear Models (GLM), and two Classification models: Random Forests (RF), and Decision Trees (RPART). The models were run on 5,10,15, and 20-fold CV, as well as at no CV for diagnostics and comparison purposes. Excel was then used to graph and record the results. The results show that Classification models are significantly better at determining the Coefficient of Determination (R2) than Regression models (P=0.01). The Root Mean Square Error (RMSE) did not significantly differ between Regression and Classification models (P= 0.96). Thus, the Classification model explained more variation while having similar RMSE to Regression. The best performing overall is the RF model at 5-fold CV, which had an R2 of 0.78. The worst performing model was the GLM at 15-Fold CV, which had the lowest R2 at 0.17. This study has revealed that increases in CV do not always improve a model's success, at least al low sample sizes. Additionally, the nonlinear nature of microbiome data advantages classification models because the do not assume linear relationships and normal distributions.

Keywords: 2026

How to Cite:

Barnes, K., Dimonaco, N., Dzinkevicius, S., Huws, S., Shirali, M. & Yan, T., (2026) “Delivering Machine Learning Approaches to predict Methane Emission Production via Microbial Features”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2286754. doi: https://doi.org/10.31274/wcgalp.24110

Rights: 1