Biology Reference
In-Depth Information
In Fig. 3(b), which shows the overall combined scores
-
, we can see
that there is a strong positive correlation between survival and
Z
-AURKA
(
0.73). Genes corresponding to the NKI and EMC signatures are
again highlighted. Also highlighted are the integrated signature genes,
defined by choosing those genes most strongly associated with survival
based on the combined Cox regression, |
-
|
ρ =
≥
6.5. It is notable that many
NKI and EMC signature genes are clustered near the middle of the sur-
vival distribution, indicating that they are not predictive of survival across
studies.
8. Conclusion
Meta-analytical approaches are valuable not only for increasing the sam-
ple size (thus decreasing sampling error artifacts), but also for reducing
the contribution of strong but nonreproducible associations caused by
platform or cohort-specific biases.
Here, we have outlined a general procedure for combining informa-
tion across independent genomic studies. Our procedure relies on the
availability and careful curation of primary data from the original studies.
The strategy produces a straightforward combination of
z
-scores, which
can be used to rank genes according to evidence for the study question
of interest. We have assumed sufficiently large sample sizes so that the
test statistics should each be approximately normally distributed. This
assumption will be satisfied for many clinical studies, but is less likely to
hold in the case of experimental studies. Studies with smaller sample sizes
may be included as well for a suitably transformed statistic.
Because we use all genes measured in each study rather than focus-
ing on only those genes found to be associated with outcome, conclu-
sions based on the combined analysis should be less subject to the
problem of publication bias. In addition, the original data handling and
analysis are less likely to impact conclusions of meta-analysis based on the
data rather than with the original data summaries.
We have also assumed that each study is independent, but it is not dif-
ficult to envision scenarios where independence is violated. For example,
it is increasingly common to obtain longitudinal data on the same set of
