Performing Department
(N/A)
Non Technical Summary
To feed the growing human population, global food production needs to increase. However, this agricultural intensification must be sustainable and protect both food security and environmental health, as agriculture is a major driver of climate change through enteric fermentation. Enteric fermentation is the process by which feed is broken down by microbes in the rumen (specialized stomach compartment of cows), which allows cattle to utilize fibrous plants as energy sources, unlike many other animals. However, enteric fermentation also results in the production of methane, a powerful greenhouse gas. Recent advances in our understanding of how this methane is produced have led to the development of substances added to an animal's feed, or Animal Feed Additives (AFA), that can reduce enteric methane emissions. However, many interacting factors impact methane production, such as cattle type, weight, feed intake, and the composition of the cow's diet. In addition, AFA can have indirect effects that may increase the excretion of other environmental pollutants by cows. To reduce methane as much as possible with AFA, we need to know how the effectiveness of AFA depends on all these interacting factors. However, this can be very challenging to assess in experiments with real animals, where it is difficult or impossible to measure all the possible interacting variables. To complement studies in real animals, animal scientists also use mathematical models, or a series of equations that describe how the cow's feed is turned into methane, to answer questions about enteric fermentation and how it may be affected by AFA.I have previously developed a comprehensive mathematical model of enteric fermentation in dairy cows that is designed to answer these questions. My model predicts how much methane a cow will produce, as well as many other variables, depending on what she is eating and which and how much AFA she is fed. However, I have not yet assessed how accurate the model's predictions are, what parts of the model are most important to making its predictions better, or how certain or uncertain the model's predictions are. In this project, my objectives are to: 1) evaluate the accuracy of the model's predictions in several different scenarios and to compare this accuracy to that of similar models; 2) conduct analyses that will tell me which parts of the model are the most important to its predictions, and 3) assess how confident we are in the model's predictions, given uncertainties in inputs to the model. After I evaluate the model with these analyses, my last objective is to use the model to find a diet and AFA combination that most effectively decreases methane emissions from dairy cows.Achieving these goals will confirm that the model is a useful, accurate tool that animal scientists and policy makers feel confident using to answer questions about the interactions between AFA use and methane production. We can then use this tool to find ways to maximize AFA efficiency, minimize greenhouse gas emissions from dairy cows, or ask more basic questions about enteric fermentation, ultimately helping us reduce the environmental impacts of animal agriculture.
Animal Health Component
0%
Research Effort Categories
Basic
100%
Applied
0%
Developmental
(N/A)
Goals / Objectives
The overall goal of this project is to evaluate a previously developed mechanistic model of rumen fermentation and use this model to optimize methane-inhibiting animal feed additive (AFA) use. This goal is divided into three objectives:Evaluate model predictive performance on independent datasetsSub aim 1a: Assess model predictive performance.Sub aim 1b: Compare model performance to similar models.Evaluate sensitivity of model to fitted parameter values, calibrate parameters, and quantify uncertainty in model outputsSub aim 2a: Conduct global sensitivity analysis.Sub aim 2b: Calibrate parameters.Sub aim 2c: Conduct uncertainty analysis.Optimize basal diet and AFA implementation scheme for maximum efficacy.
Project Methods
Aim 1: Evaluate model predictive performance on independent dataset and compare performance to related models. Sub aim 1a: Assess model predictive performance. An evaluation dataset of hourly fermentation dynamics from dairy cattle will be used to assess model predictions of diurnal enteric CH4 and H2 emissions, VFA concentrations, microbial dynamics, and rumen pH from simulations with no AFA supplementation. Data from an A. taxiformis trial in dairy cattle will be used to assess performance of daily enteric CH4 predictions from simulations including bromoform supplementation. A dataset of dairy cattle supplemented with 3NOP will include model performance as assessed by mean square prediction error (MSPE) and square root of MSPE (RMSPE) of the model's predictions from observations in each test dataset. MPSE will be decomposed into mean, slope, and random bias. The ratio of the RMSPE to the standard deviation of the observations (RSR) will also be calculated and used to compare performance across datasets.Sub aim 1b: Compare model performance to similar models. Model performance on the 3NOP dataset will be assessed for a rumen mechanistic model incorporating 3NOP using the same performance metrics as above, while model performance on the A. taxiformis dataset will be assess for a mechanistic model incorporating A. taxiformis. Evaluation milestones for Aim 1b will include error index statistics (MSPE, RMSPE, and RSR) for each model on each test dataset. To assess relative model performance, Lin's Concordance Correlation Coefficient (CCC) will calculated and decomposed into the Pearson correlation coefficient, r, which measures prediction precision, and the bias correction factor, Cb, which measures prediction accuracy.Aim 2: Evaluate sensitivity of model to fitted parameter values, calibrate parameters, and quantify uncertainty in model outputs.Sub aim 2a: Conduct global sensitivity analysis. Local sensitivity analysis (SA) explores the sensitivity of model outputs to parameters by holding all parameters except one constant, and therefore does not account for parameter correlation. Global SA (GSA) overcomes this limitation by varying all inputs at the same time, and measuring the variance attributed to an input relative to the total variance of all inputs. However, regression-based GSA approaches are only applicable if the model is linear in its parameters, which may not be known a priori. Therefore, a model-free, global, variance-based SA approach will be used. Because calculating variance-based sensitivity indices for all model parameters is computationally expensive, we will first screen all parameters using the Method of Morris. The Morris method provides the measures mu*, the mean of the elementary effects of a given parameter, and sigma*, the standard deviation of the elementary effects of the parameter. We will select all parameters whose normalized (mu*, sigma*) coordinate pair is a distance of 0.5 or greater from the origin for further analysis. Parameters not selected for further analysis will be fixed to their optimized values. We will conduct a GSA by calculating the first- and total-order Sobol' indices for each selected parameter. Sobol' indices decompose variance in the model output into first-order and higher order effects of each parameter, respectively. These indices will quantify the contribution of the variance in each parameter to the variance in model outputs due to their first and higher order effects.Sub aim 2b: Calibrate parameters. The previously accomplished parameter optimization deterministically identified parameters values that minimize model errors on fitting data, but did not estimate parameter distributions or uncertainty. We will utilize a Markov Chain Monte Carlo simulation to sample values from each parameter posterior distribution that can be used to estimate mean, variance, and 95% credible intervals for all influential parameters.Sub aim 2c: Model output uncertainty analysis. We will perform an uncertainty analysis of model outputs by executing the model many times, sampling parameters from their joint posterior distribution as generated by the MCMC simulation, and the mean and standard deviation of model outputs will be estimated, quantifying the variability in all outputs due to variability in the parameters.Aim 3: Optimize basal diet and AFA implementation scheme for maximizing AFA efficacy. An optimization algorithm will be used to optimize basal diet and AFA choices that respectively minimize enteric CH4, net GHG emissions, or nutrient excretion to the environment as predicted by the mechanistic model, given fixed AFA doses. Outcomes will include a description of optimal basal diet for each AFA and their associated environmental impacts.Interpretation of results from Aims 1-3: Aim 1 will assess the model's ability to make accurate predictions and compare its accuracy to similar models. Aim 2a will identify the most globally influential model influential parameters, which I expect to include pH- and fractional outflow-related parameters. Results of Aims 1 and 2a can be used together to identify parameters with disproportionate effects on model outputs that may cause poor model predictions. Given these results, I will reconsider model structure, particularly novel elements of outflow and pH, to further improve model performance. Aims 2b and 2c will generate likely ranges for each model prediction variable based on uncertainty in inputs, which will allow us to understand the typical variation in each outcome variable and move away from deterministic "average animal" predictions. In Aim 3, we will identify diet and AFA protocols that minimize various environmental impacts, and together, Aims 2b-c and 3 will allow us to understand the possible range of effects these diets may have.Evaluation of expected outcomes: The expected outcomes of Aim 1 will provide quantitative information on the predictive performance of the mechanistic model and therefore evaluate the hypothesis embedded in model, such as the importance of pH, rumen outflow, and thermodynamics in determining enteric CH4 emissions. Evaluation of these hypotheses is an important step in developing our understanding of the effects of AFA on enteric fermentation and consolidating various hypotheses in the scientific literature on enteric fermentation into one comprehensive rumen model. Outcomes from Aim 1 will also allow fellow researchers to potentially use the evaluated model to investigate feed additive use and related questions, adding a complementary tool to in vivo studies. Results from Aim 2 will allow improved prediction of on-farm cattle GHG emissions with associated uncertainties, improving our understanding of likely ranges of these outcomes. Results from Aim 3 will identify how to implement methane-inhibiting AFA most effectively and allow the potential use of identified optimum diets to maximize AFA efficacy on-farm. The evaluation of all three aims, together, will lead to the potential reduction of agricultural GHG emissions through optimized use of methane-inhibiting feed additives.