The Jackson Laboratory has recently launched a large-scale mutagenesis program to systematically collect novel neurological mouse mutants. Beyond the logistics associated with screening large numbers of mice the greatest challenge lies in the need to detect individual deviant mice from a large population. Currently available behavioral and physiological paradigms are typically only used to reveal group differences. Automated methods for detection of a variety of phenotypes are required.

We have developed (in conjunction with Columbus Instruments, Columbus Ohio) a comprehensive cage monitoring system (CCMS) that monitors mice in specialized individual live-in cages. The system allows automated, non-invasive, simultaneous collection of 1) total, ambulatory and rearing activity; 2) food and water consumption; and 3) oxygen and carbon dioxide concentrations.

To validate CCMS as a tool for deviant detection we collected data from 29 known mutant mice - 4 tubby (obesity), 6 mdx (dystrophic), 7 Drd3 (dopamine receptor knockout), 3 Fmr1 (fragile mental X syndrome), 4 sti (ataxic), 5 het (inner ear defect) - and 13 mice from 2 additional strains (7 A/J and 6 129SV/J) and compared them to normal C57BL/6J mice. All mutants were on C57BL/6J backgrounds (except closely related B10 for mdx) and were selected because they had subtle, non-visible phenotypes. All mice were age matched (8 wk). Data were recorded continuously for three days.

To exploit the CCMS data set, we developed a multivariate statistical algorithm for mutant detection. Briefly, the algorithm first "trains" on a data set from normal mice and then computes the Mahalanobis distance [1] between subsequent individual mice and the training set. If this distance is "too" large, the mouse is flagged as an outlier. The cutoff distance for flagging can be adjusted for a desired rate of false positives (5% for these experiments).

Initial development of the algorithm used only 3 variables (ambulation, rearing and CO2/V02 ratio) and divided each of these into light and dark period values for a total of 6 parameters. The algorithm was applied to three data sets: a "training" set including 32 C57BL6/J mice, a "control" set including 19 C57BL6/J mice, and a "test" set including the mutants and the different strains. The outlier scores for the training and control sets were approximately evenly distributed between 0 and 1 (Figure 1a,b respectively). In contrast, for the test set the outlier scores were concentrated near 0, indicating that some mutants were flagged (Figure 1c). The algorithm flagged 1/19 in the control set (~5% as predicted) while in the test set 12/29 mutant mice and 12/13 mice of a different strain (not in figure) were flagged.

These results demonstrate that CCMS can detect a variety of mutants and can provide an interpretative foundation for subsequent focused tests. Ongoing work is aimed at increasing detection by including more variables (43 summary variables are currently derived from raw CCMS data sets) and determining if subsets of these variables can be grouped to target specific phenotypes.

References

1. Krzanowski, W.J. (1988). Principles of Multivariate Analysis: A User's Perspective. Oxford: Oxford University Press.

© 2000 Noldus Information Technology b.v.