|
- |
-
|
-
- |
-
-
|
-||||
-
-
|
-||||
Open the default .htm file to view this Web presentation.
This presentation contains content that your browser is unable to display. This presentation was optimized for the recent version of Microsoft Internet Explorer and the recent version of Netscape Navigator.
| WebQTL Demonstration
One please link to www.webqtl.org/search.html |
| "or webqtl.org/search.html..." |
| Search results |
| "First page of data:" |
| "Data sources:" |
| "Return to Trait Data page" |
| "Discovering shared expression patterns" |
| "The App transcript neighborhood" |
| "Handdrawn sketch of the neighborhood" |
| "What a network is likely..." |
| Are there experimental results to corroborate a link between App with Hsp84-1? |
| "2.45 billion scatter plots" |
| "Cross-tissue type correlations" |
| "Cross-modal correlations:" |
| WebQTL link to www.webqtl.org/search.html |
| Slide 16 |
| Slide 17 |
| WebQTL to exploring upstream control |
| WebQTL to exploring upstream control. |
| The whole neighborhood is modulated! |
| Which gene is the QTL? |
| Slide 22 |
| Slide 23 |
| Tissue differences in probe performance |
| Is there known biology to link Hars2 with App? |
| WebQTL link to www.webqtl.org/search.html |
| Requirement: The gene must be polymorphic to be genetically ÒupstreamÓ |
| Direct correlations between genotypes and traits |
| WhatÕs downstream of Chr 2 near Hars2? |
| WhatÕs downstream of Chr 2 near Hars2? |
| Does Hars2 correlate with Actn2 strongly? |
| Contact for comments and improvements: |
| WebQTL Demonstration
One please link to www.webqtl.org/search.html |
| Part 1: How to discover shared expression patterns (slides 2Ð14) | |
| Part 2. Discovering upstream modulators (15Ð25) | |
| Discovering downstream targets |
| "or webqtl.org/search.html..." |
| or webqtl.org/search.html (mirror) |
| Search results |
| "First page of data:" |
| First page of data: The ÒTrait Data FormÓ |
| "Data sources:" |
| Data sources: Phenotpyes and genotypes |
| "Return to Trait Data page" |
| Return to Trait Data page |
| "Discovering shared expression patterns" |
| Discovering shared expression patterns |
| "The App transcript neighborhood" |
| The App transcript neighborhood |
| "Handdrawn sketch of the neighborhood" |
| Handdrawn sketch of the neighborhood |
| "What a network is likely..." |
| What a network is likely to look like (but App will not be center of universe). |
| Are there experimental results to corroborate a link between App with Hsp84-1? |
| "2.45 billion scatter plots" |
| 2.45 billion scatter plots: here is one of the best |
| "Cross-tissue type correlations" |
| Cross-tissue type correlations |
| "Cross-modal correlations:" |
| Cross-modal correlations: From mRNA to to anatomy and to behavior and pharmacology |
| WebQTL link to www.webqtl.org/search.html |
| Discovering shared expression patterns | |
| Discovering upstream modulators (QTLs) | |
| Discovering downstream targets |
| Slide 16 |
| Slide 17 |
| WebQTL to exploring upstream control |
| WebQTL to exploring upstream control. |
| The whole neighborhood is modulated! |
| Which gene is the QTL? |
| Slide 22 |
| Slide 23 |
| Tissue differences in probe performance |
| Is there known biology to link Hars2 with App? |
| WebQTL link to www.webqtl.org/search.html |
| Discovering shared expression patterns | |
| Discovering upstream modulators (QTLs) | |
| Discovering downstream targets |
| Requirement: The gene must be polymorphic to be genetically ÒupstreamÓ |
| Direct correlations between genotypes and traits |
| WhatÕs downstream of Chr 2 near Hars2? |
| WhatÕs downstream of Chr 2 near Hars2? |
| Does Hars2 correlate with Actn2 strongly? |
| Contact for comments and improvements: |
' + ENDSHOW_MESG + '
| Welcome to a short
demonstation of WebQTL. Please adjust the wize of windows on your monitor so
that you can see part of this page as well as WebQTL windows. WebQTL will
produce a potentially large number of new windows (pop-ups), so you may need
to modify your browser preferences to permit pop-ups. In this demonstration, we
explore one important transcript expressed in the brain: the amyloid beta
precursor protein messenger RNA. The product of this mRNA, the APP protein,
is associated with Alzheimer disease. |
|
| (Initial version of June
2003 by Rob Williams, Last edits June 16, 2003 by RW.) |
| WebQTL can be used to
enter your own trait data or to work with data that we have entered for you. |
|
| Linking to
http://www.webqtl.org/search.html will get you a version of the window above.
It may not be identical in layout but it will have the key features. Please
follow the intructions on the slide. Of course, we encourage you to enter
your own terms of interest. |
|
| Two points: If you make
a search term too complex you may get no hits. if you make it too simple (for
example, APP) then you may get too much. Experiment. |
|
| If you just link to |
|
| http://www.webqtl.org
you will NOT see the window above but will see text that will help you to
enter your own data. To get to a
version of the window shown above you will need to click on the phrase RNA expression and Phenotype
Databases in the upper banner. |
|
| If you do not get a new
page within 30 seconds then there is
a problem: try the mirror site http://webqtl.org/search.html. |
| If all goes well, you
will see a version of this window. WebQTL will display up to about 100 hits.
If a search generates larger numbers of hits then you will need to refine
your search terms. |
| The Trait Data and
Editing Form is the single more important page from the point of view of
working with WebQTL data. Please read the text carefully. Explore the links,
but do not close this page. We will need it many more times in this
demonstration. |
| There are already five
databases in WebQTL. Each will eventually have a page like this describing
the data source and appropriate citations to these databases. The great
majority of data in WebQTL were generated in our own labs and those of our
collaborators. We welcome you to
use these data, but caution you that there are inevitably lots of little
problems and issues that may compromise some results. Be cautious and
skeptical. Ask us questions before you leap to publication. And please, if
you find the data useful or can verify or refute data, LET US KNOW. We would
like to add you to our reference section and add links to improvements. |
| This slide shows you the lower parts of the Trait Data Page.
We expect to make many small modification of this page, so do not be
surprised if some elements have been moved around. |
| Finally, we can now
start an analysis. |
|
| We ask a simple
question: |
|
| Do differences in App
transcript expression correlate with those of any other transcripts in the
forebrain? |
| The answer is a strong
Yes. A very large number of transcripts have correlations above 0.7 (absolute
value) with App mRNA. The precise number today is 208. But this will change
as we add more strains and arrays. In any case, this is a fairly large number
and all of these correlations are significant at alpha .05 even when
correcting for the enormous numbers
of tests (12422 tests). |
|
| What does this imply? |
|
| That there can be
massive codependence of expression variance among transcripts. App is NOT an
isolated instance. This is improtant biologically and statistically. From a
statistical perspective, we would like to know how many ÒindependentÓ test we
effectively are performing when we use array data in this way. Are we testing
12000 independent transcripts or just 1200 transcriptional ÒmodulesÓ each
with blurred boarders but each with about 10 effective members. There is no
answer yet, but we probaby have a large enough data set to begin to answer
this question. |
| Many of the data types
in the previous slide are hot-linked and it is easy to generate a small web
of correlations between any transcript of interest and many other
transcripts. In this case, we have used green lines between transcripts that
have positive correlations, and red lines between transcripts that have
negative correlations. Correlations have been multiplied by 100. The
correlation of 0.96 between App and Hsp84-1 reads 96. These are Pearson product moment
correlations and they are sensitive to outliers. If you prefer, you can
recompute Spearman rank order correlations. |
|
| Where did Ndr4 (lower
left) come from? It is not in the list in the previous slide. Actually it is.
Nomenclature changes rapidly. If you click on R74996 in the previous slide
(the active webqtl version) you will see that it now has a new symbol and
name. |
|
| What are all of the conventions in this correlation
network sketch. |
|
| 1.The official gene symbol = App |
|
| 2. OUr estimate of the
location of these gene in the Mouse Genome Sequencing Consoritum version 3
build (MGSCv3). Chromosome followed by the megabase position relative to the
centromere. (Mice only have one chromosome arm so this is an unambiguous
coordinate. ) |
|
| 3. The pair of numbers:
top is the highest expression among the strain set. The lower number is the
lowest expression of that transcript among the strain set. |
|
| 4. Vertical number on
the right side of each box: this is the probe set ID given by Affymetrix. We
have truncated these probe set IDs so you will not see the usual ÒatÓ. A single gene may be
represented by more than 10 probe sets. Thus this ID number is essential to
identify the actual data source. |
|
| 5. Lower right corner: a
two digit number followed by plus and minus signs. These numbers are the
correlation value (absolute value) of the 100th best correlated transcript.
The plus and minus signs indicate the mean polarity of the correlations. |
|
| 6. The set of numbers
that read 2@140* etc. These are the approximate locations of additive effect
QTLs detected by WebQTL that we will describe in other slides. Read this as:
App has a suggestive QTL on Chr 2 at about 140 Mb and the D allele inherited
from DBA/2J confirms a higher expression level at this marker. If there is no star symbol, then it
is not even formally ÒsuggestiveÓ but does make an interesting looking blip
on the QTL radar screen. |
| What networks are likely
to really look like. This slide is taken from Lumeta Inc. (www.lumeta.com). It actually
illustrates the structure of connections across the Internet. The large green area is a major Internet
provider (WorldCom before the fall?).
Check out Lumeta to see
some more lovely graphs of this sort. Most biologists are familiar with
simple sketches, but this is what we will have to be prepared to contend with
soon. |
| Having worked with
WebQTL now for 30 minutes, do we know anything new? The hypothesis that we
have generated (but not validated) is that three transcripts: Atp6l, Gnas,
and Ndr4 are part of a family of genes that are coregulated in normal mouse
forebrain with App and Hsp84-1. We need to add functional and mechanistic
significance to this hypothesis to make it biologically vibrant. But from a
statiistical standpoint it is a strong inference. |
|
| Please donÕt say: But
these are mere correlations. A high correlation in this context has a
biological basis. The real question is are we smart enough to understand the
web (not chain) of causality that produced the correlation. Once we
understand the web of causality, does it have utility? Very often the answer
will be NO. This will often be the case when a high correlation is generated
by linkage disequilibrium of sets of polymorphisms that modulate a set of
mechanistically separated traits. Chromosomal linkage can produce
correlations that are not mechanistic in the conventional sense used by
molecular biologists. For example, clusters of hox transcription factor genes tend to be close
physically to keratin gene clusters, and one might expect shared patterns of
variance produced by this linkage in a mapping panel, no matter how large. |
|
| If Affymetrix designed
probe sets with reasonable care, if we did the experiments correctly, if we
sampled animals appropriately, then a correlation of 0.70 or higher between
transcripts in the brain tells you that these two transcripts are effectively
coupled in this set of animals under this set of conditions. More than 50%
the variance in the expression of one transcript can be predicted from the
other. That is a major piece of information that could be of significant
clinical, economic, and predictive value, whatever its causes. Yes,
correlation coefficients are noisy and have large error terms, but we have
larger n of strains coming to the rescue. Expect more than 50 BXD lines soon. |
|
| This is a thin veneer of
functional genomics. It is enough to generate some marvelous hypotheses in a
semi-automated way. |
| The correlation between
App and heat shock protein 84-1 transcript is most impressive. Since WebQTL now contains total of
about 70,000 traits in the BXD strains, we could produce as many as to 70k x
35k scatter plots of this type. Since all of the correlations come for a common reference population,
none of the correlations are
blantantly silly. However the great majority may be uninterpretable and a
very large number may be meaningless given the signal-to-noise ratios of some
measurements. With about 30 strains, correlations above 0.7 have a reasonably
low false positive rate. |
| We can compare App
expression inthe forebrain against transcript expression in hematopoietic
stem cells. Some of these correlations are significant, but it may be
difficult to discovery of shared genetic (linkage disequilibrium) or
molecular processes that give rise to these correlation. |
|
| The GNF Hematopoietic
stem cell data belong to Gerald de Haan (University of Groningen) and Michael
Cooke (Genomics Institute of the Novartis Research Foundation). |
| Another example, but in
this case we are generating correlations between variation in transcript
levels with a database of approximately 430 published (and unpublished)
phenotypes from BXD strains. Notice that the N of strains is variable (from
21 to 28 above). Rank order statistics is more appropriate when N is under
30. |
|
| The Published Phenotypes
database was prepared by Elissa Chesler and Robert Williams from data
extracted from the literature or sent to us for inclusion by our colleagues.
We especially thank John Crabbe (Oregon HSU) and Byron Jones (Pennsylvania
SU) for providing us with large pre-compiled data tables. |
| Part 2: Mapping upstream
modulators or QTLs. A quantitative trait locus is a chromosomal region that
harbors one or a few polymorphic gene loci that influence a trait. We are
going to be looking for QTLs that modulate the steady state expression level
of App in the adult mouse forebrain. |
| The next set of slides
provide a very short interlude on QTL mapping. You will need to do some
independent reading on this topic if this is your first exposure to QTL
mapping. The recombinant inbred strains that we are using in WebQTL and in
this particular demo were generated about 25 years ago by Dr. Ben Taylor at
The Jackson Laboratory. He crossed a female C57BL/6J mouse with a male DBA/2J
mice. At the bottom of this slide we have schematized one chromosome pair
from three out of 80 BXD RI strains.
The dashed vertical lines that lead to the final BXD RI lines involve
20 full sib matings (about 6 years of breeding). Some lines die out during inbreeding. For example,
there is no extant BXD3 or BXD4 strain. |
| This slide is
illustrates two categories of QTLs that modulate variability in transcript
abundance. |
|
| 1. cis QTLs are defined
as QTLs that are closely linked to the gene whose transcript is the measured
trait. For example, a polymorphism in the promoter that affects the binding
of a transcription factor. However, cis QTLs can be far upstream or downstream
polymorphisms in enhancers. They may also be polymorphisms in neigghboring
genes. |
|
| 2. trans QTLs map far
enough away from the location of the gene that gives rise to the transcript
that is being measured so that we can be quite certain that the QTL is not IN
the gene itself. The most blatant type of trans-QTL would be a polymorphism
in a transcription factor. BUT in the majority of cases, the trans QTLs can
be far removed in a mechanistic sense from the actual events modulating
transcript abundance. That is why there are three overlappoing arrows above. The way in which an upstream
polymorphism influences a downstream difference in mRNA abundance can be very
indirect. Effects can : |
|
| a. cross tissue types (a polymorphic liver enzyme may affect
CNS gene expression) |
|
| b. cross time (the modulator is only expressed for one day
during development but has permanent effects in adults), |
|
| c. may be contingent on environmental factors (heat shock may
trigger the expression of a polymorphic factor that affects mRNA abundance). |
| Back to the demo. Please
bring the Traiit Data and Editing window to the front and look for the
Interval Mapping button. Please confirm that you are back to the trait
amyloid beta precursor protein.
If so, then just click the button. |
|
| Notice that the default
for: |
|
| Select Chrs (chromosomes)
is ALL |
|
| Select Probes is Probe Set |
|
| Options: Permuation test
YES (1000 is the default number) |
|
| Options: Bootstrap test
YES (1000 is the default number) |
| This is the main output
type: a so-called full genome interval map. |
|
| The X-axis represents all
19 autosomes and the X chromosome as if they were laid end to end with short
gaps between the telomere of one chromosome and the centromere of the next
chromosome (mouse chromsomes only have a single long arm and the centromere represents
the origin of each chromosome for numerical purpose: 0 centimorgans and
almost 0 megabases). The blue labels along the bottom of the figure list a
subset of markers that were used in mapping. We used 753 markers to perform
the mapping but here we just list five markers per chromosome. |
|
| The thick blue wavy line
running across chromsomes summarizes the strength of association between
variation in the phenotype (App expression differences) and the two genotypes
of 753 markers and the intervals between markers (hence, interval mapping). The height of the wave (blue Y-axis
to the left) provides the likelihood ratio statistic (LRS). Divide by 4.61 to
convert these values to LOD scores.
Or you can read them as a chi-square-like statistic. |
|
| The red line and the red
axis to the far right provides an estimate of the effect that a QTL has on expression of App
(this estimate of the addtive effect tends to be an overestimate). If the red
line is below the X-axis then this means that the allele inherited from
C57BL/6J (B6 or B) at a particular marker is associated with higher values.
If the red line is above the X-axis then the DBA/2J allele (D2 or D) is
associated with higher traits. Multiply the additive effect size by 2 to
estimate the difference between the set of strains that have the B/B genotype
and the D/D genotype at a specific marker. For example, on Chr 2 the red
line peaks at a value of about 0.25. That means that this
region of chromosome 2 is responsible for a 0.5 unit expression difference
between B/B strains and the D/D strains. Since the units are log base 2, this
is 2^0.5, or about a 41% difference in expression with the D/D group being
high. |
|
| The yellow histogram bars:
These summarize the results of a whole-genome bootstrap of the trait that is
performed 1000 times. What is a bootstrap? A bootstrap provides you a metho
of evaluating whether results are robust. If we drop out one strain, do we
still get the same results? When mapping quantitative traits, each strain
normally gets one equally weighted vote. But inthe bootstrap procedure, we
give each strain a random weighting factor of between 0 and 1. We then remap the trait and find THE
SINGLE BEST LRS VALUE per bootstrap. We do this 1000 times. In this example,
most bootstrap results cluster on Chr 2 under the LRS peak. That is somewhat
reassuring. But notice that a substantial number of bootstrap results prefer
Chr 7 or Chr 18. |
|
| The horizontal dashed
lines at 9.6 and 15.9. These lines are the LRS values associated with the
suggestive and significant false positive rates for genome-wide scans
established by permuations of phenotypes across genotypes. We shuffle
randomly 1000 times and obtain a distribution of peak LRS scores to generate
a null distribution. Five percent of the time, one of these permuted data
sets will have a peak LRS higher than 15.9. We call that level the 0.05
significance threshold for a whole genome scan. The p = 0.67 point is the the
suggestive level, and corresponds to the green dashed line. These thresholds are conservative for
transcripts that have expression variation that is highly heritable. The
putative or suggestive QTL on Chr 2 is probably more than just suggestive. |
|
| One other point: the
mapping procedure we use is computationally very fast, but it is relatively
simple. We are not looking for gene-gene interactions and we are not fitting
multiple QTLs in combinations.
Consider this QTL analysis a first pass that will highlight hot spots
and warm spots that are worth following up on using more sophisticated
models. |
|
| CLICKABLE REGIONS: |
|
| 1. If you click on the
Chromosome number then you will generate a new map just for that chromosome. |
|
| 2. If you click on the
body of the map, say on the blue line, then you will generate a view on a 10 Mb window of that part of the
genome from the UCSC Genome Browser web site. |
|
| 3. If you click on a
marker symbol, then you will generate a new Trait data and editing window
with the genotypes loaded into the window just like any other trait. More on
this later. |
|
| NOTE: you can drag these
maps off of the browser window and onto your desktop. The will be saved as
PNG or PDF files. You can import them into Photoshop or other programs. |
| App has a suggestive QTL
on Chr 2. What about the neighbors that we defined as having shared
expression patterns. This figure shows that members of the immediate App
neigborhood share a weak Chr 2 QTL.
That is what the blue oval in the background is meant to represent.
But some transcripts, such as Ndr4 and Psen2 do not share this Chr 2
interval. |
|
| QUESTION: What kind of
headway can we make in detemining what polymorphism or polymorphisms on Chr 2
near 130 Mb might contribute to the variance in the expression of all of
these important transcripts? |
| Candidate
Genes: The best we can do at
this point is to make an educated guess about the candidacy status of all
genes in the QTL support interval. For sake of argument, lets say that we are
confident that the polymorphism is located between 130 and 150 Mb (20 Mb,
equivalent to roughly 10 cM). There will typically be 12 to 15 genes per Mb,
so we now would like to evaluate 240 to 300 positional candidates. We would
like to highlight the biologically relevant subset of candidates. We could
look through gene ontologies and expression levels to help us winnow the
list. An alternate way avaiable using WebQTL is to generate a list of those
genes in this 20 Mb interval that have transcripts that co-vary in expression
with App expression. |
|
| To do this, go
back to the Trait Data and Editing window. Sort the correlations by position.
Select Return = 500. Then scroll down the list to see positional candidates
that share expression with App. |
|
| There are
several candidates that have high correlation with App even in this short 20
Mb interval. We can rank them by correlation, but they are all close. There is one other imporant approach
to ranking these candidates. Are they likely to contain polymorphisms? We can
assess the likelihood that they contain polymorphisms by mapping each
transcript to see if any have strong cis QTLs. The logic of this search is
that a transcript that has a strong cis-QTL is likely to contain functional
polymorphisms that effect its own expression. This make is more like that the
transcript is a ÒcausativeÓ factor since it is likely to be polymorphic. |
|
| When you do this you
will find that only the transcript 0610006H08Rik has a strong cis QTL. See
the slide above. The LRS peaks above 35
(LOD of greater than 7.5). It turns out that this transcript is really
Hars2, also known as histydl tRNA synthase 2. |
| LetÕs look at Hars2 in
more detail by mapping all of the perfect match probes (16 of them) that
target this transcript. |
|
| Go back to the Trait
Data and Editing window and select Chr 2 (rather than ALL as shown above) and
also select PM Probes. Then click on Interval Mapping button. |
|
| You will get the
illustration above, but without the sequence data that we have added. The 16 perfect match probes are
arranged in sequence (red is 5 prime, blue is the 3 prime end). For example,
the 5 prime-most primer 307387 has the sequence CACTG..... It also has a
polymorphism at the 17 nucleotide of this 25 nt probe sequence. |
|
| How do we know that the
5 prime probe is polymorphic? By looking up the sequence in the Celera
Genomics databases which often contqains sequence data for C57BL/6J (B6
above) and for DBA/2J. But two
blue probes (14 and 15) do NOT contain SNPs but still have very large LRS
scores. The other probes do not perform so wel. Highly variable probe
performance is probably a result of the very different stacking energies of
DNA-RNA duplexes. |
| The vertical text says it
all: Even when using identical probes, mapping performance (and signal)
depends on tissue type and mRNA complexity. This is another gene in the Hars2
interval. Forebrain and tem cell mRNAs were run on the same U74Av2 array,
whereas the cerebellum mRNA was run on the 430A and 430B array set. |
| Hars2 is not a well
characterized gene and their is no biology yet to support the hypothesis that
Hars2 modulates gene expression, let alone App expression in specific. There
are also serious database/assembly discrepancies between Celera and MGSCv3
regarding the genomic organization of this gene. But there appear to be
approximately 69 SNPs in Hars2, one of which results in a substitution. |
| Part 3. Many investigators would like to
discover the set of downstream targets of a gene of interest. |
|
| In a genetic and
functional sense, that question can only be addressed effectively if there is
genetic variation in the particular gene. We know that Fos is an important transcription factor, but
unless it is polymorphic between C57BL/6J and DBA/2J, then it cannot generate
a genetic variance signal with which we can work. We can still study
covariance of Fos and hundreds of other transcripts (an interesting
exercise), but there are no genetic causes-and-effects. |
| Genes must be polymorphic
to generate downstream genetic effects (as opposed to downstream molecular
effects). Hars2 meets this condition because we have already mapped a
functional polymorphism in the gene. We can therefore posit that Hars2 is a
QT gene. What transcripts are downstream? App is one obvious
candidate, but there are many more. |
|
| The are several ways to
look for downstream targets. The best and most obvious is to look up all
transcripts that have high correlations with Hars2 itself. You should know
how to do this. An alternative method is shown here for teaching purpose and
to show you what to do if your gene of interest is not in our database. You
need to know: |
|
| 1. Where your gene is located. You need this information to find a
surrogate marker; a marker that is located very close to your gene of
interest. |
|
| 2. That your gene is
polymorphic between C57BL/6J and DBA/2J. |
|
| LetÕs look at the
correlation of Hars2 with BXD genotypes as shown in the slide above to
illustrate how to use markers as surrogate traits. |
|
| Go to the Trait Date and
Editing window one more time. We want the data for Hars2 this time, not App.
You should be able to show that Hars2 has a high correlation with D2Mit423 as shown in the slide above. |
|
| By clicking on the symbol
D2Mit423 in the Correlation window, you will generate a new Trait Data window
shown on the next slide. |
| We can review the set of
correlations between the marker D2Mit423 and all transcripts in
forebrain. This is in essence a
backwards way of mapping QTLs. We are considering one marker and asking what
traits correlate to the marker and how well. |
| The marker D2Mit423
correlates moderately well with a number of Chr 2 transcripts. This is due to
linkage disequilibrium. These correlations are not due to a molecular
interactions other than being close together on a chromosome. But we have circled one transcript,
actinin alpha 2, that has a moderately good correlation (0.59) with D2Mit423.
If we map this gene we expect to find a suggestive QTL that peaks near
D2Mit423 |
| There is some support for
the hypothesis that Actn2 is downstream of a polymorphism in the Hars2
region. But again, due to the 10 to 20 Mb precision of the mapping data, this
relation could be generated by a large number of other polymorphisms close to
Hars2. In the absence of a biological connection between Actn2 and Hars2 we
have a weak hypothesis. If there were a plausible functional connection
between the two genes, then this hypothesis could be quickly upgraded. |
| We can test the Hars2 to
Actn2 connection directly. This process weakens the putative association. We
are ready to move on and examine other candidates in the Hars2 region near
D2Mit423. Or in your case, please
start from the beginning using other genes and transcripts and tissues that
interest you more than this App-Hsp84-Hars2 example. |
|
| This concludes the first
demonstation of how to use some of the WebQTL features. Please explore.
Please also send feedback for improvements or additions to
rwilliam@nb.utmem.edu |
| END |
Open the default .htm file to view this Web presentation.
This presentation contains content that your browser is unable to display. This presentation was optimized for the recent version of Microsoft Internet Explorer and Netscape Navigator 4.
| The GeneNetwork and WebQTL :
PART 2 link to www.genenetwork.org |
| Slide 2 |
| Slide 3 |
| Discovering upstream modulatory loci |
| WebQTL searches for upstream controllers |
| Genetic versus Physical maps for App expression |
| Physical map for distal chromosome 7 |
| Evaluating candidate genes |
| Physical maps are zoomable |
| Evaluating Ctbp2 as a candidate QTL for App |
| Evaluating Ctbp2 using other resources |
| "Summary of Part 2" |
| Test Questions |
| Contact for comments and improvements: |
| The GeneNetwork and WebQTL :
PART 2 link to www.genenetwork.org |
| Part 1. How to study expression variation and genetic correlation (slides 2–17) | |
| Part 2. Discovering upstream modulators (slides 18–29) |
| Slide 2 |
| Slide 3 |
| Discovering upstream modulatory loci |
| WebQTL searches for upstream controllers |
| Genetic versus Physical maps for App expression |
| Physical map for distal chromosome 7 |
| Evaluating candidate genes |
| Physical maps are zoomable |
| Evaluating Ctbp2 as a candidate QTL for App |
| Evaluating Ctbp2 using other resources |
| "Summary of Part 2" |
| Summary of Part 2 |
| Test Questions |
| Contact for comments and improvements: |
' + ENDSHOW_MESG + '
| Part 2: Discovering
upstream modulators and quantitative trait loci (QTLs). A quantitative trait
locus is a chromosomal region that harbors one or a few polymorphic gene loci
that influence a trait. We are going to be looking for QTLs that modulate the
steady state expression level of App in the adult mouse forebrain. |
| The next few slides
provide a short introduction to mapping the loci that are responsible for
variation in a trait such as App expression level. These modulatory regions
of the genome are sometimes called quantitative trait loci or QTLs. You may
want to do some independent reading on this topic if this is your first
exposure to QTL analysis. |
|
| The genetic reference
population (GRP) of BXD recombinant inbred strains were originally generated
about 25 years ago by Benjamin Taylor at The Jackson Laboratory. He crossed
female C57BL/6J mice with male DBA/2J mice to generate the F1 and F2 progeny.
At the bottom of this slide we have schematized one chromosome pair from
three of the BXD RI strains. The
dashed vertical lines that lead to the final BXD RI lines involve 21 full sib
matings (about 7 years of breeding). Some lines die out during inbreeding.
For example, there is no longer any BXD3 strain. |
|
| Notes: |
|
| 1. Over the last decade,
our group (Lu Lu and Rob Williams) and Jeremy Peirce and Lee Silver at
Princeton have enlarged Ben TaylorÕs set. There are now just over 80 BXD
strains. They have all been genotyped using about 13,700 markers (SNPs and
microsatellites). These markers are used to define the ÒblueÓ and ÒredÓ
regions of the chromosomes as shown in the figure above. |
|
| 2. Chromosomes of RI GRPs
usually have about 4 times as many recombinations as those of F2 animals.
However, unlike an F2, both chromosomes of an RI are identical. Therefore, 50
RI strains contain as many recombinations as 100 F2 animals. |
|
| 3. BXD43 through BXD100
were generated using a special method that resulted in a further doubling of
the average recombination density per chromosome. The entire set of 80 BXDs
therefore contains as many recombinations as about 260 F2 animals. |
| This slide is
illustrates two major types of QTLs that modulate variability in
transcript-relative steady state abundance. |
|
| 1. cis QTLs are defined
as QTLs that are closely linked to the gene whose transcript is the measured
trait. For example, a polymorphism in the promoter that affects binding of a
transcription factor. However, cis QTLs can be far upstream or downstream polymorphisms
in enhancers or may be in 3Õ UTR binding sites that affect message stability. |
|
| 2. trans QTLs map far
enough away from the location of the gene that gives rise to the transcript
that is being measured so that we can be fairly certain that the QTL is not
in the gene itself. The most blatant type of trans QTL would be a
polymorphism in a transcription factor. But in the majority of cases, the
trans QTLs can be far removed in a mechanistic sense from the actual events
modulating transcript abundance. That is why there are three overlapping
arrows in the figure. The way in
which an upstream polymorphism influences a downstream difference in mRNA
abundance can be indirect. Effects can: |
|
| a. cross tissue types (a polymorphic liver enzyme may affect
CNS gene expression) |
|
| b. cross time (the modulator is only expressed for one day
during development but has permanent effects in adults) |
|
| c. may be contingent on environmental factors (heat shock may
trigger the expression of a polymorphic factor that affects mRNA abundance). |
| Please bring the Trait
Data and Analysis window to the front and look for the Interval Mapping
button. Confirm that you are back to the trait amyloid beta precursor
protein. If so, then just click
the button. |
|
| Notice that the default
for: |
|
| Select Chrs (chromosomes)
is ALL |
|
| Select Mapping Scale is
set to GENETIC |
|
| Options: Permutation test
YES (2000 is the default number) |
|
| Options: Bootstrap test
YES (2000 is the default number) |
| This is a major output
type: a so-called full-genome interval map. |
|
| The X-axis represents all
19 autosomes and the X chromosome as if they were laid end to end with short
gaps between the telomere of one chromosome and the centromere of the next
chromosome (mouse chromosomes only have a single long arm and the centromere
represents the origin of each chromosome for numerical purpose: 0
centimorgans at almost 0 megabases). The blue labels along the bottom of the
figure list a subset of the 3795 markers that were used in mapping. |
|
| The thick blue wavy line
running across chromosomes summarizes the strength of association between
variation in the phenotype (App expression differences) and the two genotypes
of all markers and the intervals between markers (hence, interval mapping). The height of the wave (blue Y-axis
to the left) provides the likelihood ratio statistic (LRS). Divide by 4.61 to
convert these values to LOD scores.
Or you can read them as a chi-square-like statistic. |
|
| The red line and the red
axis to the far right provide an estimate of the effect that a QTL has on
expression of App (this estimate of the so-called additive effect tends to be
too high). If the red line is below the X-axis then this means that the allele
inherited from C57BL/6J (B6 or B) at a particular marker is associated with
higher values. If the red line is above the X-axis then the DBA/2J allele (D2
or D) is associated with higher trait values. Multiply the additive effect
size by 2 to estimate the difference between the set of strains that have the
B/B genotype and those that have the D/D genotype at a specific marker. For
example, on distal Chr 7 the red line peaks at a value of about 0.2. That
means that this region of chromosome 2 is responsible for a 0.4 unit
expression difference between B/B strains and the D/D strains. |
|
| The yellow histogram bars:
These summarize the results of a whole-genome bootstrap of the trait that is
performed 1000 times. What is a bootstrap? A bootstrap provides a method to
evaluate whether results are robust. If we drop out one strain, do we still
get the same results? When mapping quantitative traits, each strain normally
gets one equally weighted vote. But using the bootstrap procedure, we give
each strain a random weighting factor of between 0 and 1. We then remap the trait and find THE
SINGLE BEST LRS VALUE per bootstrap. We do this 1000 times. In this example,
most bootstrap results cluster on Chr 3 and Chr 7 under the LRS peaks. That
is somewhat reassuring. But notice that a substantial number of bootstrap are
scattered around on other chromosomes. About 30% of the bootstrap resamples
have a peak on Chr 7. That is pretty good, but does makes us realize that the
sample we are working with is still quite small and fragile. |
|
| The horizontal dashed
lines at 10.5 and 17.3 are the likelihood ratio statistic (LRS) values
associated with the suggestive and significant genome-wide probabilities that
were established by permutations of phenotypes across genotypes. We shuffle
randomly 2000 times and obtain a distribution of peak LRS scores to generate
a null distribution. Five percent of the time, one of these permuted data
sets will have a peak LRS higher than 17.3. We call that level the 0.05
significance threshold for a whole genome scan. The p = 0.67 point is the
suggestive level, and corresponds to the green dashed line. These thresholds are conservative for
transcripts that have expression variation that is highly heritable. The putative
or suggestive QTL on Chr 3 is probably more than just suggestive. |
|
| One other point: the
mapping procedure we use is computationally very fast, but it is relatively
simple. We are not looking for gene-gene interactions and we are not fitting
multiple QTLs in combinations. Consider this QTL analysis a first pass that
will highlight hot spots and warm spots that are worth following up on using
more sophisticated models. |
|
| CLICKABLE REGIONS: |
|
| 1. If you click on the
Chromosome number then you will generate a new map just for that chromosome. |
|
| 2. If you click on the
body of the map, say on the blue line, then you will generate a view on a 10
Mb window of that part of the genome from the UCSC Genome Browser web site. |
|
| 3. If you click on a
marker symbol, then you will generate a new Trait data and Analysis window
with the genotypes loaded into the window just like any other trait. More on
this in Section 3. |
|
| 4. You can drag these maps
off of the browser window and onto your desktop. They will be saved as PNG or
PDF files. You can import them into Photoshop or other programs. |
|
| 5. There is also an option
at the bottom of the page to download a 2X higher resolution image of this
plot for papers and presentations. |
|
| 6. You can also download
the results of the analysis in a text format |
| The map on the top has an
X-axis scale based on frequency of recombinations events between markers (B
to D transitions, see slide 19 for a color-coded example). These so-called
genetic maps are scaled in centimorgan (recombinations per 100 gametes). In contrast,
the physical map shown below the genetic map has an X-axis scale based on DNA
length measured in nucleotides or base-pairs. Notice the large difference
between the two maps in the size of Chr 19 (large on the genetic scale but
small on the physical scale). |
|
| Also notice the large
difference in the width of the chromosome 7 QTL peak. In mice, recombinations
occur with higher frequency toward the telomeric side (right side) of each
chromosome. As a result, genetic maps are stretched out more toward the
telomere relative to a physical map. The QTL on distal Chr 7 is therefore
actually more precisely mapped than might appear looking at the genetic map. |
|
| The physical scale is
becoming more useful than the genetic scale primarily because many other data
types can be easily superimposed on a physical map. You will see more
examples in the next several slides. |
| Physical map of variation
in App expression in brain on distal Chr 7 (a blow up of the whole-genome map
on the previous slide). |
|
| Notes: |
|
| 1. You can now see that the X-axis is on a physical scale of megabases
(Mb). The QTL peak is roughly between 120 and 132 Mb. |
|
| 2. The small irregular
colored blocks and marks toward the top of the map mark the locations of
genes superimposed on the physical map. Neighboring genes are offset slightly
in the vertical axis for display purpose. Note one region of very high gene
density from about 120 to 123 Mb. |
|
| 3. The orange hash marks
along the X-axis represent the number of single nucleotide polymorphisms that
distinguish the two parental strains (C57BL/6J and DBA/2J) from each other.
We call this the SNP seismograph track (see Glossary for more details). Regions
with low numbers of SNPs have closely matched sequences and are less likely
to contain QTLs. |
|
| 4. As before, the thin red
line shows the additive effect size. By convention the positive values
signify the D alleles are associated with higher expression of App in this
region of Chr 7 than the B alleles. The maximum effect size is about +0.20
log2 expression units per D allele. The differences been the BB and DD
genotypes (BB and DD because each strain has two alleles; one per chromosome)
is therefore about 2^0.4 = 1.32 or a 32% increment in DD relative to BB at
this locus. |
|
| 5. If you scroll just
under the Physical Map you will see text that reads ÒDISPLAY from XXX Mb TO
YYY MbÉ..Ó These physical maps
are zoomable, a feature we will exploit to evaluate candidate genes in this
QTL interval. |
| Evaluating
candidate genes (CHECKED BOXES) responsible for variability in APP
expression: |
|
| A large number
of genes are usually in the QTL interval and are therefore POSITIONAL
CANDIDATES, but they will differ greatly in their biological and
bioinformatic plausibility. Assume that the QTL has been located between 119
and 131 Mb (12 Mb). There will typically be 12 to 15 genes per Mb, so we
might need to evaluate several hundred positional candidates. In this
particular case there are about 100 known genes in this interval. Eight of
these are highlighted in the table above with check marks in the boxes to the
left. We need to highlight and
objectively score the biologically relevant subset of all 100 positional
candidate genes. We could look through gene ontologies and expression levels
to help us shorten the list. An alternate way available using WebQTL is to
generate a list of those genes in this interval that have transcripts that
co-vary in expression with App expression. That is what the table shows. |
|
| Notes: |
|
| 1. To replicate
this table go back to the Trait Data and Analysis Form. Choose to sort
correlations by POSITION and select RETURN = 500. Then scroll down the list
to Chr 7 and review the subset of positional candidates that share expression
with App. You should see a list similar to that shown above. Gtf3c1 is a good
biological candidate and has a high covariation in expression with App. |
|
| 2. Caveat: Of course, the gene or genes
that control App expression may not be in this list. A protein coding
difference might be the ultimate cause of variation in App transcript level
and the expression covariation might be close to zero. Our list may also
simply be missing the right transcript since the microarray is not truly
comprehensive. Furthermore, even if the list contains the QTL gene, an
expression difference may only have been expressed early in development or
even in another tissue such as liver. While it is important to recognize
these caveats, it is equally important to devise a rational way to rank
candidates given existing data. Coexpression is one of several criteria used
to evaluate positional candidates. We will see others in the next slide. |
|
| 3. We can also
assess the likelihood that candidates contain functional polymorphism in
promoters and enhancers that affect their expression simply by mapping the
transcripts of all candidate genes to see if they Òmap backÓ to the location
of gene itself. A transcript that maps to its own location is referred to as
a cis QTL. We essentially ask: Which of the transcripts listed in the
Correlation Table above (from Gtf3c1 to Zranb1) has variation in expression
that maps to Chr 7 at about 120 Mb?
The logic of this search is that if a gene controls the level of its
own expression it is also much more likely to generate other downstream
effects. The Gtf3c1 transcript is a weak cis QTL with a local LRS maximum of
about 7.0 (D alleles are high). That is just about sufficient to declare it
to be a cis QTL. [No whole genome correction is required and a point-wise
p-value of 0.05 is the appropriate test. A p-value of 0.05 is roughly
equivalent to an LRS of 6.0 (LOD = 1.3).] |
|
| An even higher blow-up of
part of the Chr 7 physical map of variation in App expression in brain. The QTL region actually extends from
about 119 to 129. |
|
| Notes: |
|
| 1. As mentioned
in the previous slide another important approach to ranking candidates is
based on the number of sequence variants that distinguish the parental
strains. If we were sure that the sequences of the gene, its promoter, and
its enhancers were identical between the strains then we could discount--but
not eliminate--that gene as a candidate. The Gtf3c1 candidate almost falls
into this category: of 663 known SNPs in and around this gene, only four
differ between C57BL/6J and DBA/2J. Gtf3c1 is essentially
identical-by-descent in these strains and is a less likely candidate. In
contrast, if the two alleles of the gene have dozens of functional variants
in exons, promoters, enhancers, and splice sites, then it becomes a higher
priority candidate. |
|
| Of course it
only takes a single critical sequence variant to generate downstream effects.
The argument above is really about the prior probabilities. Where would you
place your bets given the information at hand? |
|
| 2. If you scroll down the INTERVAL
ANALYST you will find that Ctbp2 is a particularly interesting candidate that
contains lots of SNPs (n = 75 and a SNP density of 0.55 SNP/Kb). Ctbp2 is
also closer to our QTL peak than was Gtf3c1. Not only does Ctbp2 contain lots
of SNPs but it is also is associated with a powerful cis QTL with an LRS of
24.2 (divide by 4.61 to get the equivalent LOD score of 5.25). |
|
| 3. At this high magnification,
individual genes are distinct. They are color coded by their density of SNPs.
Bright orange represents those genes that have a high SNP density (C57BL/6J
versus DBA/2J), black represents genes with low SNP density. Roll the cursor
over a gene block and its name will pop up, along with information on exon
number. |
|
| 4. Beneath the physical map you will
find an INTERVAL ANALYST table that lists information on known genes in the
region on which you have zoomed the Physical Map. |
|
| 5. As always: error-checking is
important. Some genes may be missing from the Interval Analyst (recent
additions or errors of omission). In this case the Zranb1 gene that is
located just proximal to Ctbp2 is not listed in the INTERVAL ANALYST.
Double-check the interval using the Genome Browser links (blue and beige
horizontal bars) at the top of the PHYSICAL MAP. |
| This slide illustrates one
reason why Ctbp2 should be considered a high priority positional candidate
gene that may modulate the expression level of App. Ctbp2 is a strong cis QTL in some brain regions (here the
data are taken from the striatum).
If Ctbp2 contains variants that modulate its own expression then these
expression differences may produce many downstream effects. Of course, we now
want to know much more about the known biology of Ctbp2. What kind of gene is
it? To begin to answer that question we can use a number of resources listed
in the LINKS page. |
|
| Notes: |
|
| 1. The App QTL is bimodal. Perhaps there are actually two causal factors
in this region--one close to 123 Mb and the other close to 127 Mb. |
|
| 2. The precision of QTL
mapping depends on several factors, including the effect size and
interactions among QTLs modulating a trait, the number of genetic individuals
that are studied, and the distribution of recombinations in the study
population. In the case above,
the QTL(s) are likely to be confined to the interval from 120 to 132 Mb. The
bootstrap test (yellow bars shown in some of the previous slides) can be
usual for estimating the consistency of QTL peaks. |
| Ctbp2 should also be
considered a high priority biological candidate gene responsible for
modulating App expression levels. The C-terminal binding protein 2 is a
transcriptional co-repressor also known as Ribeye. The gene produces two transcripts encoding distinct
proteins. The short form is a transcriptional repressor that binds a
Pro-X-Asp-Leu-Ser peptide motif and interacts with several transcription
factors including EVI1, ZFPM1, and ZFHX1A (aka TCF8, deltaEF1). The longer
isoform is a major component of specialized synapses in photoreceptors. Both
proteins contain a NAD+ binding domain similar to NAD+-dependent
2-hydroxyacid dehydrogenases. |
|
| Notes: |
|
| 1. To find out more about
CTBP2 protein and the Ctbp2 gene, link to iHOP at
http://www.pdg.cnb.uam.es/UniPub/iHOP/ and type in CTBP2 |
|
| Try Arrowsmith at
http://arrowsmith.psych.uic.edu/cgi-test/arrowsmith_uic/pubsmith.cgi |
|
| 2. Both APP and CTBP2 are
involved in oxidoreducatase activity or Notch signaling. To establish this
common gene ontology visit NCBI
http://www.ncbi.nih.gov/entrez/query.fcgi?db=gene and enter each gene
symbol. |
|
| 3. You can get interesting
hints regarding Ctbp2 expression partners by examining the genetic
correlations between Ctbp2 probe set 1422887_a_at and all other transcripts
on the M430 Affymetrix array. Use the Striatum data set because we already
know from previous work (the previous slide) that this gene is a cis
QTL. You should be able to show
that Ctbp2 and Notch3 have antagonistic expression patterns in striatum. The
negative genetic correlation with E2f4 is even stronger. The transcript also
has a high positive genetic correlation with Rdh14. Of particular interest
with respect to APP protein processing, Ctbp2 covaries positively with Bace2
(the transcript of the beta site APP-cleaving enzyme 2). |
|
| By clicking on the
CORRELATION of the Atcay transcript to the App transcript, you can generate a
Correlation plot between these two transcripts. In this App and Atcay
scatterplot, each point is a strain mean value. For example, BXD33 and BXD8
have low App and Atcay expressions. The two parental strains and the F1 are
also included in this plot. |
| A group of traits from
many different databases can be selected and brought together for joint
analysis. In this case all of the content of the BXD SELECTIONS is from a
single BRAIN database, the top 20 neighbors of the App transcript from the
Correlation Results table. Eight of these neighbors plus App is shown in the
slide. |
|
| Notes: |
|
| 1.All of items in the BXD SELECTIONS were selected using
the SELECT ALL button |
|
| 2. The buttons at the
top (and bottom) of this page can do some cool stuff. We will work with
NETWORK GRAPH first. |
|
| 3. Think of the
SELECTIONS as your shopping cart. You go to different aisles in the
supermarket to acquire different types of items of interest. These could
include transcripts, classical phenotypes (longevity, brain weight, prepulse
inhibition, iron levels in midbrain). ÒChecking outÓ in this case involves
doing some analysis with the items in the cart. |
|
| 4. Different tools
handle different numbers of items. Most will handle up to 100 traits. |
|
| END |
| Going back to the Trait
Data and Analysis Form window, we have computed the correlations between
strain variation in App expression level and other classical phenotypes that
have already been measured in many of the same BXD strains. |
|
| Notes: |
|
| 1.The number of common strains varies widely--in this
case from 14 to 23 strains. |
|
| 2. We can add these
traits (four are selected) to our BXD SELECTIONS window. |
| We have computed the
Network Graph, now using other types of traits. |
|
| Saline Hot Plate Latency
is the green node labeled 10020. |
|
| Freezing (fear) is the
green node labeled 10447. |
|
| Notes: |
| Part 2: Discovering
upstream modulators and quantitative trait loci (QTLs). A quantitative trait
locus is a chromosomal region that harbors one or a few polymorphic gene loci
that influence a trait. We are going to be looking for QTLs that modulate the
steady state expression level of App in the adult mouse forebrain. |
| The next few slides
provide a short introduction to mapping the loci that are responsible for
variation in a trait such as App expression level. These modulatory regions
of the genome are sometimes called quantitative trait loci or QTLs. You may
want to do some independent reading on this topic if this is your first
exposure to QTL analysis. |
|
| The genetic reference
population (GRP) of BXD recombinant inbred strains were originally generated
about 25 years ago by Benjamin Taylor at The Jackson Laboratory. He crossed
female C57BL/6J mice with male DBA/2J mice to generate the F1 and F2 progeny.
At the bottom of this slide we have schematized one chromosome pair from
three of the BXD RI strains. The
dashed vertical lines that lead to the final BXD RI lines involve 21 full sib
matings (about 7 years of breeding). Some lines die out during inbreeding.
For example, there is no longer any BXD3 strain. |
|
| Notes: |
|
| 1. Over the last decade,
our group (Lu Lu and Rob Williams) and Jeremy Peirce and Lee Silver at
Princeton have enlarged Ben TaylorÕs set. There are now just over 80 BXD
strains. They have all been genotyped using about 13,700 markers (SNPs and
microsatellites). These markers are used to define the ÒblueÓ and ÒredÓ
regions of the chromosomes as shown in the figure above. |
|
| 2. Chromosomes of RI GRPs
usually have about 4 times as many recombinations as those of F2 animals.
However, unlike an F2, both chromosomes of an RI are identical. Therefore, 50
RI strains contain as many recombinations as 100 F2 animals. |
|
| 3. BXD43 through BXD100
were generated using a special method that resulted in a further doubling of
the average recombination density per chromosome. The entire set of 80 BXDs
therefore contains as many recombinations as about 260 F2 animals. |
| This slide is
illustrates two major types of QTLs that modulate variability in
transcript-relative steady state abundance. |
|
| 1. cis QTLs are defined
as QTLs that are closely linked to the gene whose transcript is the measured
trait. For example, a polymorphism in the promoter that affects binding of a
transcription factor. However, cis QTLs can be far upstream or downstream polymorphisms
in enhancers or may be in 3Õ UTR binding sites that affect message stability. |
|
| 2. trans QTLs map far
enough away from the location of the gene that gives rise to the transcript
that is being measured so that we can be fairly certain that the QTL is not
in the gene itself. The most blatant type of trans QTL would be a
polymorphism in a transcription factor. But in the majority of cases, the
trans QTLs can be far removed in a mechanistic sense from the actual events
modulating transcript abundance. That is why there are three overlapping
arrows in the figure. The way in
which an upstream polymorphism influences a downstream difference in mRNA
abundance can be indirect. Effects can: |
|
| a. cross tissue types (a polymorphic liver enzyme may affect
CNS gene expression) |
|
| b. cross time (the modulator is only expressed for one day
during development but has permanent effects in adults) |
|
| c. may be contingent on environmental factors (heat shock may
trigger the expression of a polymorphic factor that affects mRNA abundance). |
| Please bring the Trait
Data and Analysis window to the front and look for the Interval Mapping
button. Confirm that you are back to the trait amyloid beta precursor
protein. If so, then just click
the button. |
|
| Notice that the default
for: |
|
| Select Chrs (chromosomes)
is ALL |
|
| Select Mapping Scale is
set to GENETIC |
|
| Options: Permutation test
YES (2000 is the default number) |
|
| Options: Bootstrap test
YES (2000 is the default number) |
| This is a major output
type: a so-called full-genome interval map. |
|
| The X-axis represents all
19 autosomes and the X chromosome as if they were laid end to end with short
gaps between the telomere of one chromosome and the centromere of the next
chromosome (mouse chromosomes only have a single long arm and the centromere
represents the origin of each chromosome for numerical purpose: 0
centimorgans at almost 0 megabases). The blue labels along the bottom of the
figure list a subset of the 3795 markers that were used in mapping. |
|
| The thick blue wavy line
running across chromosomes summarizes the strength of association between
variation in the phenotype (App expression differences) and the two genotypes
of all markers and the intervals between markers (hence, interval mapping). The height of the wave (blue Y-axis
to the left) provides the likelihood ratio statistic (LRS). Divide by 4.61 to
convert these values to LOD scores.
Or you can read them as a chi-square-like statistic. |
|
| The red line and the red
axis to the far right provides an estimate of the effect that a QTL has on
expression of App (this estimate of the so-called additive effect tends to be
too high). If the red line is below the X-axis then this means that the allele
inherited from C57BL/6J (B6 or B) at a particular marker is associated with
higher values. If the red line is above the X-axis then the DBA/2J allele (D2
or D) is associated with higher trait values. Multiply the additive effect
size by 2 to estimate the difference between the set of strains that have the
B/B genotype and those that have the D/D genotype at a specific marker. For
example, on distal Chr 7 the red line peaks at a value of about 0.2. That
means that this region of chromosome 2 is responsible for a 0.4 unit
expression difference between B/B strains and the D/D strains. |
|
| The yellow histogram bars:
These summarize the results of a whole-genome bootstrap of the trait that is
performed 1000 times. What is a bootstrap? A bootstrap provides a method to
evaluate whether results are robust. If we drop out one strain, do we still
get the same results? When mapping quantitative traits, each strain normally
gets one equally weighted vote. But using the bootstrap procedure, we give
each strain a random weighting factor of between 0 and 1. We then remap the trait and find THE
SINGLE BEST LRS VALUE per bootstrap. We do this 1000 times. In this example,
most bootstrap results cluster on Chr 3 and Chr 7 under the LRS peaks. That
is somewhat reassuring. But notice that a substantial number of bootstrap are
scattered around on other chromosomes. About 30% of the bootstrap resamples
have a peak on Chr 7. That is pretty good, but does makes us realize that the
sample we are working with is still quite small and fragile. |
|
| The horizontal dashed
lines at 10.5 and 17.3 are the likelihood ratio statistic (LRS) values
associated with the suggestive and significant genome-wide probabilities that
were established by permutations of phenotypes across genotypes. We shuffle
randomly 2000 times and obtain a distribution of peak LRS scores to generate
a null distribution. Five percent of the time, one of these permuted data
sets will have a peak LRS higher than 17.3. We call that level the 0.05
significance threshold for a whole genome scan. The p = 0.67 point is the the
suggestive level, and corresponds to the green dashed line. These thresholds are conservative for
transcripts that have expression variation that is highly heritable. The
putative or suggestive QTL on Chr 3 is probably more than just suggestive. |
|
| One other point: the
mapping procedure we use is computationally very fast, but it is relatively
simple. We are not looking for gene-gene interactions and we are not fitting
multiple QTLs in combinations. Consider this QTL analysis a first pass that
will highlight hot spots and warm spots that are worth following up on using
more sophisticated models. |
|
| CLICKABLE REGIONS: |
|
| 1. If you click on the
Chromosome number then you will generate a new map just for that chromosome. |
|
| 2. If you click on the
body of the map, say on the blue line, then you will generate a view on a 10
Mb window of that part of the genome from the UCSC Genome Browser web site. |
|
| 3. If you click on a
marker symbol, then you will generate a new Trait data and Analysis window
with the genotypes loaded into the window just like any other trait. More on
this in Section 3. |
|
| 4. You can drag these maps
off of the browser window and onto your desktop. They will be saved as PNG or
PDF files. You can import them into Photoshop or other programs. |
|
| 5. There is also an option
at the bottom of the page to download a 2X higher resolution image of this
plot for papers and presentations. |
|
| 6. You can also download
the results of the analysis in a text format |
| The map on the top has an
X-axis scale based on frequency of recombinations events between markers (B
to D transitions, see slide 19 for a color-coded example). These so-called
genetic maps are scaled in centimorgan (recombinations per 100 gametes). In contrast,
the physical map shown below the genetic map has an X-axis scale based on DNA
length measured in nucleotides or base-pairs. Notice the large difference
between the two maps in the size of Chr 19 (large on the genetic scale but
small on the physical scale). |
|
| Also notice the large
difference in the width of the chromosome 7 QTL peak. In mice, recombinations
occur with higher frequency toward the telomeric side (righ sidet) of each
chromosome. As a result, genetic maps are stretched out more toward the
telomere relative to a physical map. The QTL on distal Chr 7 is therefore
actually more precisely mapped than might appear looking at the genetic map. |
|
| The physical scale is
becoming more useful than the genetic scale primarily because many other data
types can be easily superimposed on a physical map. You will see more
examples in the next several slides. |
| Physical map of variation
in App expression in brain on distal Chr 7 (a blow up of the whole-genome map
on the previous slide). |
|
| Notes: |
|
| 1. You can now see that the X-axis is on a physical scale of megabases
(Mb). The QTL peak is roughly between 120 and 132 Mb. |
|
| 2. The small irregular
colored blocks and marks toward the top of the map mark the locations of
genes superimposed on the physical map. Neighboring genes are offset slightly
in the vertical axis for display purpose. Note one region of very high gene
density from about 120 to 123 Mb. |
|
| 3. The orange hash marks
along the X-axis represent the number of single nucleotide polymorphisms that
distinguish the two parental strains (C57BL/6J and DBA/2J) from each other.
We call this the SNP seismograph track (see Glossary for more details). Regions
with low numbers of SNP have closely matched sequences and are less likely to
contain QTLs. |
|
| 4. As before, the thin red
line shows the additive effect size. By convention the positive values
signify the D alleles are associated with higher expression of App in this
region of Chr 7 than the B alleles. The maximum effect size is about +0.20
log2 expression units per D allele. The differences been the BB and DD
genotypes (BB and DD because each strain has two alleles; one per chromosome)
is therefore about 2^0.4 = 1.32; or a 32% increment in DD relative to BB at
this locus. |
|
| 5. If you scroll just
under the Physical Map you will see text that reads ÒDISPLAY from XXX Mb TO
YYY MbÉ..Ó These physical maps
are zoomable, a feature we will exploit to evaluate candidate genes in this
QTL interval. |
| Evaluating
candidate genes (CHECKED BOXES) responsible for variability in APP
expression: |
|
| A large number
of genes are usually in the QTL interval and are therefore POSITIONAL
CANDIDATES, but they will differ greatly in their biological and
bioinformatic plausibility. Assume that the QTL has been located between 119
and 131 Mb (12 Mb). There will typically be 12 to 15 genes per Mb, so we
might need to evaluate several hundred positional candidates. In this
particular case there are about 100 known genes in this interval. Eight of
these are highlighted in the table above with check marks in the boxes to the
left. We need to highlight and
objectively score the biologically relevant subset of all 100 positional
candidate genes. We could look through gene ontologies and expression levels
to help us shorten the list. An alternate way available using WebQTL is to
generate a list of those genes in this interval that have transcripts that
co-vary in expression with App expression. That is what the table shows. |
|
| Notes: |
|
| 1. To replicate
this table go back to the Trait Data and Analysis Form. Choose to sort
correlations by POSITION and select RETURN = 500. Then scroll down the list
to Chr 7 and review the subset of positional candidates that share expression
with App. You should see a list similar to that shown above. Gtf3c1 is a good
biological candidate and has a high covariation in expression with App. |
|
| 2. Caveat: Of course, the gene or genes
that control App expression may not be in this list. A protein coding
difference might be the ultimate cause of variation in App transcript level
and the expression covariation might be close to zero. Our list may also
simply be missing the right transcript since the microarray is not truly
comprehensive. Furthermore, even if the list contains the QT gene, an
expression difference may only have been expressed early in development or
even in another tissue such as liver. While it is important to recognize
these caveats, it is equally important to devise a rational way to rank
candidates given existing data. Coexpression is one of several criteria used
to evaluate positional candidates. We will see others in the next slide. |
|
| 3. We can also
assess the likelihood that candidates contain functional polymorphism in
promoters and enhancers that affect their expression simply by mapping the
transcripts of all candidate genes to see if they Òmap backÓ to the location
of gene itself. A transcript that maps to its own location is referred to as
a cis QTL. We essentially ask: Which of the the transcripts listed in the
Correlation Table above (from Gtf3c1 to Zranb1) has variation in expression
that maps to Chr 7 at about 120 Mb?
The logic of this search is that if a gene controls the level of its
own expression it is also much more likely to generate other downstream
effects. The Gtf3c1 transcript is a weak cis QTL with a local LRS maximum of
about 7.0 (D alleles are high). That is just about sufficient to declare it
to be a cis QTL. [No whole genome correction is required and a point-wise
p-value of 0.05 is the appropriate test. A p-value of 0.05 is roughly
equivalent to an LRS of 6.0 (LOD = 1.3).] |
|
| Even higher blow-up of
part of the Chr 7 physical map of variation in App expression in brain. The QTL region actually extends from
about 119 to 129. |
|
| Notes: |
|
| 1. As mentioned
in the previous slide another important approach to ranking candidates is
based on the number of sequence variants that distinguish the parental
strains. If we were sure that the sequences of the gene, its promoter, and
its enhancers were identical between the strains then we could discount--but
not eliminate--that gene as a candidate. The Gtf3c1 candidate almost falls
into this category: of 663 known SNPs in and around this gene, only four
differ between C57BL/6J and DBA/2J. Gtf3c1 is essentially
identical-by-descent in these strains and is a less likely candidate. In
contrast, if the two alleles of the gene have dozens of functional variants
in exons, promoters, enhancers, and splice sites, then it becomes a higher
priority candidate. |
|
| Of course it
only takes a single critical sequence variant to generate downstream effects.
The argument above is really about the prior probabilities. Where would you
place your bets given the information at hand? |
|
| 2. If you scroll down the INTERVAL
ANALYST you will find that Ctbp2 is a particularly interesting candidate that
contains lots of SNPs (n = 75 and a SNP density of 0.55 SNP/Kb). Ctbp2 is
also closer to our QTL peak than was Gtf3c1. Not only does Ctbp2 contain lots
of SNPs but it is also is associated with a powerful cis QTL with an LRS of
24.2 (divide by 4.61 to get the equivalent LOD score of 5.25). |
|
| 3. At this high magnification,
individual genes are distinct. They are color coded by their density of SNPs.
Bright orange represents those genes that have a high SNP density (C57BL/6J
versus DBA/2J), black represents genes with low SNP density. Roll the cursor
over a gene block and its name will pop up, along with information on exon
number. |
|
| 4. Beneath the physical map you will
find an INTERVAL ANALYST table that lists information on known genes in the
region on which you have zoomed the Physical Map. |
|
| 5. As always: error-checking is
important. Some genes may be missing from the Interval Analyst (recent
additions or errors of omission). In this case the Zranb1 gene that is
located just proximal to Ctbp2 is not listed in the INTERVAL ANALYST.
Double-check the interval using the Genome Browser links (blue and beige
horizontal bars) at the top of the PHYSICAL MAP. |
| This slide illustrates one
reason why Ctbp2 should be considered a high priority positional candidate
gene that may modulate the expression level of App. Ctbp2 is a strong cis QTL in some brain regions (here the
data are taken from the striatum).
If Ctbp2 contains variants that modulate its own expression then these
expression differencess may produce many downstream effects. Of course, we
now want to know much more about the known biology of Ctbp2. What kind of
gene is it? To begin to answer that question we can use a number of resources
listed in the LINKS page. |
|
| Notes: |
|
| 1. The App QTL is bimodal. Perhaps there are actually two causal factors
in this region--one close to 123 Mb and the other close to 127 Mb. |
|
| 2. The precision of QTL
mapping depends on several factors, including the effect size and
interactions among QTLs modulating a trait, the number of genetic individuals
that are studied, and the distribution of recombinations in the study
population. In the case above,
the QTL(s) are likely to be confined to the interval from 120 to 132 Mb. The
bootstrap test (yellow bars shown in some of the previous slides) can be
usual for estimating the consiistency of QTL peaks. |
| Ctbp2 should also be
considered a high priority biological candidate gene responsible for
modulating App expression levels. The
C-terminal binding protein 2 is a transcriptional co-repressor also
known as Ribeye. The gene produces two
transcripts encoding distinct proteins. The short form is a transcriptional
repressor that binds a Pro-X-Asp-Leu-Ser peptide motif common to adenoviral
oncoprotein E1a and a related motif in BKLF. This short form also interacts with several transcription factors
including EVI1, ZFPM1, and
ZFHX1A (aka TCF8, deltaEF1). The longer isoform is a major component
of specialized synapses in photoreceptors. Both proteins contain a NAD+
binding domain similar to NAD+-dependent 2-hydroxyacid dehydrogenases. |
|
| Notes: |
|
| 1. To find out more about
CTBP2 protein and the Ctbp2 gene, link to iHOP at
http://www.pdg.cnb.uam.es/UniPub/iHOP/ and type in CTBP2 |
|
| Try Arrowsmith at
http://arrowsmith.psych.uic.edu/cgi-test/arrowsmith_uic/pubsmith.cgi |
|
| 2. Both APP and CTBP2 are
involved in oxidoreducatase activity or Notch signalling. To estabilish this
common gene ontology visit NCBI
http://www.ncbi.nih.gov/entrez/query.fcgi?db=gene and enter each gene symbol. |
|
| 3. You can get intersting
hints regarding Ctbp2 expression partners by examining the genetic
correlations between Ctbp2 probe set 1422887_a_at and all other transcripts
on the M430 Affymetrix array. Use the Striatum data set because we already
know from previous work (the previous slide) that this gene is a cis
QTL. You should be able to show
that Ctbp2 and Notch3 have antagonistic expression patterns in striatum. The
negative genetic correlation with E2f4 is even stronger. The transcript also
has a high positive genetic correlation with Rdh14. Of particualr interest
with respect to APP protein processing, Ctbp2 covaries positiviely with Bace2
(the transcript of the beta site APP-cleaving enzyme 2). |
|
| END |
Open the default .htm file to view this Web presentation.
This presentation contains content that your browser is unable to display. This presentation was optimized for the recent version of Microsoft Internet Explorer and Netscape Navigator 4.
| GeneNetwork and WebQTL: |
| Slide 2 |
| Slide 3 |
| Search results |
| "First page of data:" |
| "Data sources:" |
| "Expression estimates for App on..." |
| "Critiquing the App data the..." |
| "App expression after windsorizing" |
| "Discovering shared expression patterns" |
| "Transcript neighborhoods" |
| "App and Atcay transcript scatterplot" |
| "App transcript and eight of..." |
| App transcript coexpression neighborhood |
| "Correlations of App with classical..." |
| "Network Graph of App with..." |
| "Summary of Part 1:" |
| Contact for comments and improvements: |
| GeneNetwork and WebQTL: |
| Part 1: How to study expression variation and covariation (slides 2–16) | |
| Part 2. Discovering upstream modulators (slides 17–30) |
| Slide 2 |
| Slide 3 |
| Search results |
| "First page of data:" |
| First page of data: The Trait Data and Analysis Form |
| "Data sources:" |
| Data sources: Metadata for each resource |
| "Expression estimates for App on..." |
| Expression estimates for App on the Trait Data form |
| "Critiquing the App data the..." |
| Critiquing the App data the Trait Data |
| "App expression after windsorizing" |
| App expression after windsorizing |
| "Discovering shared expression patterns" |
| Discovering shared expression patterns |
| "Transcript neighborhoods" |
| Transcript neighborhoods |
| "App and Atcay transcript scatterplot" |
| App and Atcay transcript scatterplot |
| "App transcript and eight of..." |
| App transcript and eight of its neighbors |
| App transcript coexpression neighborhood |
| "Correlations of App with classical..." |
| Correlations of App with classical traits |
| "Network Graph of App with..." |
| Network Graph of App with classical traits |
| "Summary of Part 1:" |
| Summary of Part 1: |
| Contact for comments and improvements: |
' + ENDSHOW_MESG + '
| Welcome to a short
demonstration of the GeneNetwork and its WebQTL module. Please adjust the
size of the windows on your monitor so that you can see part of this page, as
well as GeneNetwork windows. WebQTL produces a large number of new windows,
so you may need to modify your browser preferences to permit pop-ups. In this
demonstration, we explore one important transcript expressed in the brain:
the amyloid beta precursor protein messenger RNA. A protein product of this
mRNA, the APP protein, is associated with AlzheimerÕs disease. |
|
| My thanks to Dr. Robert
F. Clark and Wenli Cai for testing this PowerPoint demonstration and making
many improvements. |
|
| (Initial version of June
2003 by Rob Williams. Edits July 13, 2005 by RW and RFC. Edit July 14, 2005
by WC. Final edits by RF Clark, July 22, 2005. |
| Please link to the web
site: http://www.genenetwork.org |
|
| To begin a search you
make choices about what species, group, and database to explore. |
|
| For this demonstration
enter APP as above and click on the SEARCH button. Make sure that the DEFAULT
SETTINGS are species = Mouse, Group = BXD, Type = Whole Brain, and Database =
INIA BRAIN mRNA M430 (Apr05) PDNN. |
|
| Notes: |
|
| 1. The GeneNetwork and
WebQTL are often used to work with public data sets. However, it is possible
to enter and analyze your own data for specific genetic reference populations
such as the BXD genetic reference population of mice or the HXB strains of rat.
Entering your own data is a more advanced topic, but if you click on the HOME
pop-down menu (upper left), you will see ÒEnter Trait DataÓ that will explain
the process. |
|
| 2. For help on advanced searching methods read the left
side of the page (INTRODUCTION).
If you make a search term too complex, you may get no hits (try
entering Òamyloid betaÓ for example). If you make it too simple, you may also
get too many. |
|
| 3. Use the asterisk * as a wildcard. For example, to find
all Hoxb transcripts, search for Hoxb*. |
|
| 4. In some cases you can also research for transcripts
and genes using special search strings such as ÒMb = (Chr1, 100 102)Ó to find all genes on Chromosome
1 between 98 and 104 megabases (donÕt actually use the quotes). Details are
described at http://www.genenetwork.org/searchHelp.html. |
|
| 5. These INFO buttons provide
links to data about the different data types. Try them. |
|
| 6. The SET TO DEFAULT button is used to
change the database default setting to match your typical search categories. |
| SHORT DETOUR to the HELP
menu. If you are new to the GeneNetwork, you may find it helpful to review
the The Glossary and FAQ pages shown above. We are in the process of making
ÒliveÓ demos for some of the key modules in the GeneNetwork. Check the NEWS
every month or two to find out about new features. |
|
| RESULTS OF THE APP
SEARCH. A search of the INIA
Brain database generates 18 matches, 10 of which are shown above. The
GeneNetwork will display several hundred matches in pages of 40 each. If a
search generates a larger numbers of hits, then you will need to refine
search terms. |
|
| Notes: |
|
| 1. APP is a great
transcript to introduce you to the complexity and power of new array
platforms that often provide ÒalternativeÓ expression estimates for single
genes. There are seven probe sets that target different parts of the APP
transcript. Which of the alternative measurements is most appropriate and
informative? Have a look at the FAQ page for more on this topic, but general
advice: 1. be skeptical and try to validate that the correct transcript and
gene is being measured; 2. check what part of the transcript is complementary
to the probes; 3. evaluate the performance of individual probes based on
expression level, signal-to-noise and other error terms such as the standard
deviation and error. |
|
| 2. In this particular
case we have highlighted and selected # 5 on this SEARCH RESULTS page. The
annotation for this probe set mentions that it targets the last three exons
and the 3Õ untranslated region (UTR) of the amyloid precursor protein (APP). That is just what we want. |
|
| 3. Most probe sets have
not been annotated in as much detail as App. Refer tot the FAQ to learn how
to annotate probe sets yourself. |
| The Trait Data and
Analysis Form is the single most important page from the point of view of
working with GeneNetwork data. Please read the text carefully. Explore the
links, but do not close this page. We will need it many more times in this
demonstration. |
|
| Notes: |
|
| 1. What is this
database? It is called INIA Brain mRNA M430 (Apr05) PDNN, but what does that
actually mean. How much of the brain was used? How were the animals
processed? Most of these types of questions can be answered by clicking on
the DATABASE link. |
|
| 2. Transcript/gene
LOCATION data is usually from the most recent assembly. You can VERIFY the
location of the probes and probe set using the two VERIFY buttons. VERIFY
UCSC performs a sequence alignment (BLAT analysis) of the probes to the most
recent assembly. |
|
| 3. The PROBE TOOL button
provides you with highly detailed information on the probe sequences used to
assemble the probe set. For example, in this case you can find out which
probes correspond to which of the three exons. You can also review the
performance of the individual probes. Please check the GLOSSARY for
additional details on probes. |
|
| 4. The identifiers (IDs)
provide links to other key resources. |
|
| Most of the database
components and resources in The GeneNetwork are linked to metadata pages that
provide a human-readable summary of how, why, where, when, and with whom the
data were generated. Before you get too involved with a data set, it is
naturally important to read this information. While the data in The
GeneNetwork may be accessible and useful, that does not always mean that the
data is public domain and available for you to use in publication or for
profit purposes. If you want to know more about the data ownership and usage,
please read through the POLICIES pop-down menu items. |
| This slide shows you the
lower parts of the Trait Data and Analysis Form with the data for the first
set of BXD strains |
| The Basic Statistics
page allows you to evaluate some aspects of the data quality. In this case,
BXD8 is a potential problem. An outlier of this type may be generated by a
technical artifact (bad sample?). However, it is also possible that BXD8 just
has genuine low endogenous expression of App and may therefore be a
particularly valuable model for research. There are different ways to treat
problematic data of these types. One way is simply to discard this datum. The
other way is to prevent outliers from have too much influence quantitatively,
while leaving them in their low (or high positions). This is called
windsorizing the data (after King Henry the VIII who had a habit of chopping
heads). In this case, we have windsorized the BXD8 to a value of 16.0 and the
BXD33 to a value of 16.02. Rank is retained. We are making a bet that the two
lowest strains are really low, but we are hedging our bet and just making
them a little lower than BXD90. This removes their ÒundueÓ influence. |
|
| Notes: |
|
| 1. It turns out that
BXD8 is a strain with many odd phenotypes. The whole strain is essentially an
outlier for many traits. Therefore, the low App expression data may be quite
accurate. Still, it would be comforting to have at least two more replicates. |
| Now we get a much better
feel for the variation in the error among the cases. Those without error bars
are of course the ÒnoisiestÓ of all. This data set is not complete yet (the
aim is to acquire at least one male-female sample for each BXD strain). |
| Finally, we can now
start an analysis. |
|
| We ask a simple
question: |
|
| Do differences in this
particular App transcript steady-state abundance level correlate with those
of any other transcripts in the same INIA Brain mRNA M430 data set? |
|
| Notes: |
|
| 1.You can CHOOSE many other DATABASES at this point if
you want, but for now letÕs stick with the default. |
|
| 2. There are different
ways to sort the correlations. The most obvious is by p-value (most
significant values at the top of the list), but it is also interesting to
sort the top 100 or top 500 by their gene symbol (gene ID) or by their
chromosomal location (position). |
|
| 3. If you donÕt want
your analysis to be sensitive to outliers, then you may want to choose to use
the Spearman Rank Order method of calculating correlations. |
| The Traits Correlation
output window (Correlation Results) compares App expression data with all
other traits in this INIA Brain data set. The most significant 100 or 500
transcripts are sorted by their p-values. The top correlation is that of the
probe set to itself (often a value of 1.0, but in this case we modified the
App values manually by windsorizing the data). The next best correlation is
to another App probe set. The fourth correlation is interesting and suggests
that there may be a link between App and a particular type of ataxia (Atcay). |
|
| Notes: |
|
| 1.Use the checkboxes to the far left to select traits
that you want to study together. Once you have selected interesting traits,
click on the ADD SELECTION button. This puts all of the selected traits into
a SELECTIONS WINDOW for other types of analysis. |
|
| 2. The p-value is not
corrected for multiple tests. A conservative approach for array data would be
to assume 10,000 nominally independent tests. Subtract 4 from the exponent
and if the value is still smaller than 0.05 you may have a real correlation. |
|
| 3. The LITERATURE
CORRELATION is a data type generated by Drs. Ramin Homayouni and Michael
Berry. Click on the header column by the asterisk for more information on
this highly useful data type. |
|
| 4. We are using Pearson
product moment correlations rather that the Spearman rank order correlation.
But you can select either in the previous step. |
| By clicking on the
CORRELATION of the Atcay transcript to the App transcript, you can generate a
Correlation plot between these two transcripts. In this App and Atcay
scatterplot, each point is a strain mean value. For example, BXD33 and BXD8
have low App and Atcay expressions. The two parental strains and the F1 are
also included in this plot. |
| A group of traits from
many different databases can be selected and brought together for joint
analysis. In this case all of the content of the BXD SELECTIONS is from a
single BRAIN database, the top 20 neighbors of the App transcript from the
Correlation Results table. Eight of these neighbors plus App is shown in the
slide. |
|
| Notes: |
|
| 1.All of items in the BXD SELECTIONS were selected using
the SELECT ALL button |
|
| 2. The buttons at the
top (and bottom) of this page can do some cool stuff. We will work with
NETWORK GRAPH first. |
|
| 3. Think of the
SELECTIONS as your shopping cart. You go to different aisles in the
supermarket to acquire different types of items of interest. These could
include transcripts, classical phenotypes (longevity, brain weight, prepulse
inhibition, iron levels in midbrain). ÒChecking outÓ in this case involves
doing some analysis with the items in the cart. |
|
| 4. Different tools
handle different numbers of items. Most will handle up to 100 traits. |
|
| Output of the Network
Graph. Warm colors (orange and red) are positive correlations above 0.5
whereas cool colors (green and blue) are negative correlations. Notes: |
|
| 1. All of the nodes (gene/transcripts) on this graph are
clickable. |
|
| 2. For this graph the
App expression values have ÒrevertedÓ to their pre-Windsorized values. |
|
| 3. To generate this
graph, we used the default setting:
Size of 16 by 16 inches; Gene Symbols; Don't Show Correlations; Use
curved lines (aka ÒedgesÓ). |
| Going back to the Trait
Data and Analysis Form window, we have computed the correlations between
strain variation in App expression level and other classical phenotypes that
have already been measured in many of the same BXD strains. |
|
| Notes: |
|
| 1.The number of common strains varies widely--in this
case from 14 to 23 strains. |
|
| 2. We can add these
traits (four are selected) to our BXD SELECTIONS window. |
| We have computed the
Network Graph, now using other types of traits. |
|
| Saline Hot Plate Latency
is the green node labeled 10020. |
|
| Freezing (fear) is the
green node labeled 10447. |
|
| Notes: |
| END |
| END |
-
|
- |
|
- |
-
