Q-Vals (and False Discovery Rates) Made Easy

Download Report

Transcript Q-Vals (and False Discovery Rates) Made Easy

Q-Vals (and False Discovery
Rates) Made Easy
Dennis Shasha
Based on the paper
"Statistical significance for genomewide studies"
by John Storey and Robert Tibshirani
PNAS August 5, 2003 9440-9445
Challenge
• You test plants/patients/… in two settings
(or from different populations).
• You want to know which genes are
differentially expressed (alternate)
• You don’t want to make too many
mistakes (declaring a gene to be alternate
when in fact it’s null – not differentially
expressed).
First Idea
• You take p-vals of the differences in
expression.
• P-val(g) is the probability that if g is null, it
would have a difference at least this large.
• You choose a cutoff, say 0.05.
• You say all genes that differ with p-val <=
0.05 are truly different.
• What’s the problem?
Thought Experiment
• Suppose that no genes are truly
differentially expressed.
• You will conclude that about 5% of those
you called significant really are.
• Your false discovery rate (number null
among those predicted to be
alternate/number predicted to be alternate)
= 100%.
• Bad.
A Fundamental Insight
• All truly null genes (i.e. not truly
differentially expressed) are equally likely
to have any p-val.
• That is by construction of p-val: under the
null hypothesis, 1% of the genes will be in
the top 1 percentile, 1% will be in
percentile between 89 and 90th and so on.
P-val is just a way of saying percentile in
null condition.
What Do We Do With That?
• Mixture model: imagine null genes as light
blue marbles and truly different genes as
red ones.
• If the assay is decent, red marbles should
be concentrated at the low p-values.
0 …. Pval …………………………………………………1
Method We Can Use
• We don’t of course know the colors of the
marbles/we don’t know which genes are
true alternates.
• However, we know that null marbles are
equally likely to have any p-value.
• So, at the p-value where the height of the
marbles levels off, we have primarily light
blue marbles/null genes.
• Why?
Flat region starts here
Level of flat
region
0 …. Pval …………………………………………………1
Answer
• Because if all genes/marbles were null,
the heights would be about uniform.
• Provided the reds are concentrated near
the low p-vals, the flat regions will be
primarily light blues.
Example: all null
• Consider the all null case.
• All marbles are light blue.
• False discovery rate in region to left of flat
region is estimated number of white
marbles (based on flat region)/number of
marbles to left of flat region.
• This will be close to 100%
Flat region starts here
Level of flat
region
0 …. Pval …………………………………………………1
Example: all non-null
• Consider the all non-null case.
• All marbles are red and they are highly
skewed.
• Flat region is essentially zero.
• False discovery rate in region to left of flat
region is estimated number of white
marbles (based on flat region)/number of
marbles to left of flat region.
• This will be close to 0.
Flat region starts here
0 …. Pval …………………………………………………1
Example: mixed case
• Get a distribution of p-values.
• Find flat region.
• Estimate number of nulls in the left-of-flat
region by extending the flat line.
• This gives the false discovery rate.
Number of genes having pval
Possible p-value
threshold
Flat line; base level of
nulls
0 …. Pval ……………………………………………1
Example: mixed case
• What would you estimate the false
discovery rate to be in the case that we
declare the entire area to the left of the
possible p-value threshold to be
significant?
• 10%, 25%, 50%?
Number of genes having pval
Possible p-value
threshold
Flat line; base level of
nulls
0 …. Pval ……………………………………………1
Obtaining q-values from False
Discovery Rate
• Suppose we order genes from least pvalue to greatest.
• That corresponds to one of these
cartesian graphs.
• The q-value of a gene having p-value p is
exactly the False Discovery Rate if the
declared significance region had a
threshold of p.
Number of genes having pval
Q-value of a gene
having this p-val
is the FDR if this
is the significance
threshold.
Flat line; base level of
nulls
0 …. Pval ……………………………………………1
Lessons for Research
• Mushy p-values (large error bars/few
replicates) may force us to the far left in
order to get a low False Discovery Rate.
• This may eliminate genes of interest.
• If testing out a gene is not too expensive,
then we can accept a higher False
Discovery Rate – nothing magical about
0.01.
Number of genes having pval
Better p-values
avoid loss of
genes, for small
FalseDiscovery
Rate.
Flat line; base level of
nulls
0 …. Pval ……………………………………………1