Naomi Altman Penn State 2015 Dagstuhl Workshop Some topics that might be interesting Feature matching across samples and platforms Preprocessing number of features gtgt number of samples ID: 534967
Download Presentation The PPT/PDF document "Some statistical musings" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Some statistical musings
Naomi Altman
Penn State
2015
Dagstuhl
WorkshopSlide2
Some topics that might be interesting
Feature matching across
samples and platforms
Preprocessing
number of features >> number of samples
feature screening
replication and possibly other design issues
PCA and relatives
mixture modeling Slide3
Feature Matching
e.g. (simple) should we match RNA-
seq
with a gene expression microarray by “gene” or by “
oligo
”
?
protein MS with
RNA-
seq
or
ribo
-Seq
how should we match features such as methylation sites, protein binding regions, SNPs, transcripts and proteins? Slide4
Preprocessing
These plots show the concordance of 3 normalizations of the same
Affymetrix
microarray.
Dozens of methods are available for each platform.
Matching features across platforms is going to be very dependent on which set of normalizations are selected.Slide5
p
>>
n
When the number of features > number of samples:
correlations of magnitude very close to 1 are common
we can always obtain a multiple “
perfect”predictors
so selecting “interesting” features is difficult
“extreme”
p
-values,
Bayes
factors, etc become common
singular matrices occur in optimization algorithmsSlide6
p
>>
n
New statistical methods for feature selection such as “sparse” and “sure screening” selectors may be useful.
The idea of “
sure screening”
selectors is that prescreening brings us to
p
<n-1.
But … we have some high probability that all the “important” features are selected (along with others which we will screen out later).Slide7
Experimental Design
Randomization, replication and matching enhance our ability to reproduce research
In particular, replication ensures the results are not sample specific while blocking allows variability in the samples without swamping the effects
Multi-
omics
is best done on single samples measured on multiple platforms
Technical replication is seldom worth the cost compared to taking more biological replicatesSlide8
Dimension Reduction
PCA (or SVD) have many relatives that can be used to reduce the number of features using projections onto a lower dimensional space
The components are often not interpretable.
Many variations are available from both the machine learning and statistics communities.
Machine learning stresses fitting the data.
Statistics stresses fitting the data generating process.Slide9
Mixture Modeling
In many cases we can think of a sample as a mixture of subpopulations
We can use the EM algorithm or Bayesian methods to
deconvolve
into
the components.Slide10
Some other statistical topics already mentioned
missing features (present but not detected) which differ between samples
mis
-identified features
do
p
-values (or FDR estimates) matter?
multiple times; multiple cells; multiple individuals
biological variation
vs
measurement noise & error propagation
how can be enhance reproducibility (statistical issues)
can we fit complex models? should we?
the data are too big for most statistically trained folks
how are we going to train the current and next generation?