[BioC] LimmaGUI Spot Quality
smyth at wehi.edu.au
Sun Jul 4 01:19:49 CEST 2004
Many thanks for this further info.
I am taking from your remarks on AmpCh1 and AmpCh2 that we can read columns
into R and ignore the various ratio columns as these can be re-computed
from AmpCh1 and AmpCh2.
You are describing the "confidence estimate" as as an intuitive measure. I
understand the need for something intuitive. Unfortunately for use in
numerical calculations we need a measure which is quantitatively related to
something, e.g., is quantitatively related to the estimated variance of the
log-ratio is some way.
At 12:45 AM 4/07/2004, Graham Snudden wrote:
>To pick up the points raised in the mail below.
>1. The confidence estimate to which Liz refers is derived from the posterior
>distributions returned by the Bayesian framework that we are using to
>estimate the biological signal at each spot location. The underlying
>framework is relatively complex and provides a number of metrics relating to
>the signal in each channel. In order to simplify these metrics and make them
>intuitive to the end user (biologists) we generate a single confidence
>estimate. This estimate reflects the distribution of the ratio, i.e. how
>confident are we in the value calculated for the ratio. In most cases a
>tight signal distribution in each channel will lead to a high confidence
>however it is possible that a very broad distribution in one channel - a
>weak, or saturated, spot - and a very tight distribution in the other will
>also lead to a low confidence. A positive control, with near zero signal in
>one channel, will therefore return a low confidence reflecting the high
>degree of ambiguity in the actual value returned for the ratio. The
>associated confidence flag is derived from the confidence estimate by a
>simple lookup table which is under user control. This is described on the
>website if you follow the 'colour coded confidence flags' link on the
>product page; http://www.cambridgebluegnome.com/products/index.htm.
>2. The AmpCh1 and AmpCh2 columns return our estimate of the total signal in
>each channel. Clearly as we are not thresholding out an area of signal we
>have no concept of mean or median pixel intensity neither do we need to
>perform background subtraction as the amount of signal per spot is returned
>by the underlying models independent of any noise processes. The ratio is
>the ratio between the two channels.
>If you need additional technical information I could put you in touch with
>our academic founders out of the signal processing lab here in Cambridge.
>Clearly we are using the very different approach to more traditional
>threshold/template based solutions however our experience is that the
>Bayesian approach offers significant advantages in terms of robustness,
>automation, detection, accuracy and, as described above, confidence
>From: Gordon Smyth [mailto:smyth at wehi.edu.au]
>Sent: 02 July 2004 23:33
>To: Elizabeth Brooke-Powell
>Cc: 'James Wettenhall'; bioconductor at stat.math.ethz.ch;
>info at cambridgebluegnome.com
>Subject: RE: [BioC] LimmaGUI Spot Quality
>At 11:13 PM 2/07/2004, Elizabeth Brooke-Powell wrote:
> >Hi James,
> >The confidence values are give in numbers as decimals with 1 = 100%
> >confident (e.g. confidence value = 0.78) this is a value determined using
> >Bayesian statistics and is a measure of how confident the package is that
> >the spot it found is real. The package itself (BlueFuse only currently
> >available in the UK) uses a Bayesian model to iteratively find spots
> >looking. I don't know much more as it's protected, and I'm a biologist.
> >Basically I am asking if the model can take account of these numbers and
> >adjust the model appropriately.
>The answer is yes, in principle, but not without knowing how BlueFuse's
>"confidence value" is defined and what it means. Is the confidence value a
>probability? If so, of what? Is it a weight or an inverse variance? If so,
>of what? How does "confidence value" interact with the FLAG column included
>in BlueFuse output? You might not be able to answer these questions
>yourself but the BlueFuse developers can. I have not been able to find
>technical information on the BlueFuse www site sufficient to answer these
>Without knowing anything further, I would be inclined to treat the
>"confidence values" directly as weights in limma normalization and
>differential expression analyses. This is simple to do in principle, but it
>is not clear now to read the data in. The BlueFuse format is different to
>that of other two color image analysis programs. Is the RATIO column in the
>BlueFuse output the same as AMPCH1 divided by AMPCH2? If not, what are
>AMPCH1 and AMPCH2? We need to know this.
> > I am not sure in this case that pretending
> >to have GenePix will work as the numbers are not a simple 0 or 1 (good or
> >bad). If I was to try this, do I need to format the txt file of data to
> >like a GenePix file?
> >Thanks for you help,
>Dr Gordon K Smyth, Senior Research Scientist, Bioinformatics,
>Walter and Eliza Hall Institute of Medical Research,
>1G Royal Parade, Parkville, Vic 3050, Australia
>Tel: (03) 9345 2326, Fax (03) 9347 0852,
>Email: smyth at wehi.edu.au, www: http://www.statsci.org
More information about the Bioconductor