Robust and sensitive analysis of mouse knockout phenotypes

Abstract

A significant challenge of in-vivo studies is the identification of phenotypes with a method that is robust and reliable. The challenge arises from practical issues that lead to experimental designs which are not ideal. Breeding issues, particularly in the presence of fertility or fecundity problems, frequently lead to data being collected in multiple batches. This problem is acute in high throughput phenotyping programs. In addition, in a high throughput environment operational issues lead to controls not being measured on the same day as knockouts. We highlight how application of traditional methods, such as a Student's t-Test or a 2-way ANOVA, in these situations give flawed results and should not be used. We explore the use of mixed models using worked examples from Sanger Mouse Genome Project focusing on Dual-Energy X-Ray Absorptiometry data for the analysis of mouse knockout data and compare to a reference range approach. We show that mixed model analysis is more sensitive and less prone to artefacts allowing the discovery of subtle quantitative phenotypes essential for correlating a gene's function to human disease. We demonstrate how a mixed model approach has the additional advantage of being able to include covariates, such as body weight, to separate effect of genotype from these covariates. This is a particular issue in knockout studies, where body weight is a common phenotype and will enhance the precision of assigning phenotypes and the subsequent selection of lines for secondary phenotyping. The use of mixed models with in-vivo studies has value not only in improving the quality and sensitivity of the data analysis but also ethically as a method suitable for small batches which reduces the breeding burden of a colony. This will reduce the use of animals, increase throughput, and decrease cost whilst improving the quality and depth of knowledge gained.

Figure 1. An overview of the mixed method methodology implemented.

The process can be summarised as a top down methodology involving six steps to build a mixed model to query the phenotyping data.

Figure 2. Examining control data to assess batch to batch variation.

Representative time course plot showing the batch to batch variation in control data for male mice from a B6Brd;B6N-Tyr^c-Brd genetic background. Example shown is the variation seen in the fat mass variable measured in grams. For each day, data was collected a box plot is drawn as a five point summary indicating the minimum, 1^st quartile, median, 3^rd quartile and maximum. The global median fat mass value is shown with a black solid line.

Figure 3. Assessing distribution of batch for a variable.

Representative Normal Q-Q plots of the distribution of the mean for a batch in control data for male mice from a B6Brd;B6N-Tyr^c-Brd genetic background. A: weight, B: bone mineral density and C: lean mass.

Figure 4. Examining control data to assess variation in standard deviation with batch.

Representative time course plot showing the variation in standard deviation with batch in control data for mice from a B6Brd;B6N-Tyr^c-Brd genetic background. Example shown is the variation seen in the lean mass variable for male mice measured in grams. The global median is shown with a black solid line, and the 95% confidence interval is shown with dotted lines.