[BioC] lumi, Illumina Methylation 450k, and robust methylation calls

[zoppoli pietro]

Tim Rayner <tfrayner at ...> writes:

> 
> Hi,
> 
> I have recently started to use the lumi package to analyse some
> Illumina Human Methylation 450k data, and I have run into some
> problems which seem to revolve around division by zero in the
> gammaFitEM() function. I have adjusted the colour balance and quantile
> normalised as suggested in the vignette, but when I call gammaFitEM()
> the function complains (see the end of this email for a session dump).
> 
> I've traced the likely cause of the error to zero values returned by
> these calls in gammaFitEM():
> 
>         f1 <- dgamma(x - s[1], shape = k[1], scale = theta[1])
>         f2 <- dgamma(s[2] - x, shape = k[2], scale = theta[2])
> 
> The problem is that the z1 variable subsequently contains divisions by
> these zero values:
> 
>         z1 <- p[1] * f1/(p[1] * f1 + p[2] * f2)
> 
> When z1 is later used in calls to sum() in many places, this obviously
> returns NaN which causes the function to raise an exception. I think
> I've got around this by editing the function and putting na.rm=TRUE in
> each of the relevant calls to sum(), and the generated plots look
> quite believable, but I can't be sure if that's actually a valid
> approach. Is there a better way to address this problem?
> 
> Many thanks,
> 
> Tim
> 
Hi Tim,

I have your same problem even if with the Illumina Human Methylation 27k.
You wrote about editing  any sum() function with na.rm=TRUE.
It really worked for you ?
I also made the same changes but then R says that did not find the
".initialGammaEstimation function"
Such function is internal to the gammaFitEM package ...
How you solved the problem ?

In the end the results are valid or really we can't use this function at all ?

Thanks,
Pietro