Gregg et al in the Dulac lab found unexpectedly many (> 1300) imprinted genes genome-wide. However, DeVeale, der Kooy and Babak repeated that study and suggested that most of the positive calls by Gregg et al were false. These authors estimated only 175 imprinted genes genome-wide. (A side note: neither of these studies investigated the within-gene variation of imprinting across individuals.)
DeVeale et al explain the biased estimate of Gregg et al by unaccounted variation from both technical and biological sources. But by strictly adhering to the “Gregg-Dulac” definition of imprinted gene, neither DeVeale et al consider how the definition per se may impact statistical inference. This article is centered on that point.
Various definitions of imprinting
Gregg et al themselves make the crucial statement (Introduction, 2nd paragraph, of main text):
Imprinting refers to functional differences between the maternal and paternal chromosomes or alleles […], and is also used more strictly to define complete allele-specific silencing […]
This means two imprinted gene sets: a larger set defined by some (possibly small) functional difference and a smaller subset of only those genes for which the difference is very large. In their next sentence Gregg et al also describe these two sets as “biased” or “all-or-none” allelic expression, respectively.
Although Gregg et al verbally allude to these two (or more than two) sets of imprinted genes, in my understanding (detailed in the next section) they use a single mathematical definition of imprinting. This is based on their implicitly assumed1 parameter, the expected of fraction of maternal transcripts, denoted here as . Our favored definition of imprinting (or monoallelic expression) is also based on but in a characteristically different way:
| not imprinted (biallelic) | imprinted (monoallelic) | |
|---|---|---|
| Gregg-Dulac | ||
| DeVeale et al | ||
| our work |
The above definitions of imprinting allow testing specifically for what Gregg et al call maternal bias. For paternal bias the definition modify to (Gregg-Dulac) and to (our work). It is clear that the Gregg-Dulac definition of imprinting is very liberal and inclusive whereas ours is conservative and exclusive (excludes all values of between 1/2 and 0.9).
Hypothesis tests
Neither Gregg et al nor DeVeale et al gave a definition as formal as those in the above table. I deduced those formal definitions from their methodological descriptions. I summarize their methods here not only to present the deduction of their definitions but also to establish a link between their chi-squared test and my proposed test of monoallelic expression.
Minor points: ambiguities, contradictions
F1i and F1r together or separately?
The statement (Gregg SOM 2.2 p4)
In the current study, male and female samples were treated as replicates by summing the male and female F1i and F1r read data for each SNP.
…seems to contradict with (Gregg main text)…
Imprinting was assessed by tests in both initial and reciprocal crosses as described in the SOM.
Along the same line: Fig 2B clearly shows separate values for F1i and F1r. But Fig 2A is ambiguously annotated with “F1i and F1r value cutoff”, wherein the word “and” may mean that F1i and F1r was analyzed either together or separately.
“Two tailed probability”
-
DeVeale et al is even more implicit in this regard ↩