Difference between revisions of "GMM"
| Line 23: | Line 23: | ||
| − | For | + | For '''[http://www.mathworks.com/ Matlab] users''', download the source code and use the callCNVs.m script. |
| − | + | '''Users without Matlab''', can use the compiled version and the Matlab Component Runtime (MCR). (Please note, we are only providing a compiled Linux x86_64 version for now). | |
Then you can use (and edit according to your need) the shell script called "run_CallCNVs.sh" | Then you can use (and edit according to your need) the shell script called "run_CallCNVs.sh" | ||
Revision as of 13:01, 23 December 2009
Deletion, insertion and duplication events giving rise to copy number variations (CNVs) have been found genome-wide in the humans and other species.
Such genomic aberrations were identified already more than a decade ago using array-based comparative hybridization. They can also be detected using
data from SNP genotyping arrays, typically by combining the intensities of the two probes for a given SNP and comparing to the same SNP from other arrays (thus deriving a copy number ratio).
Significant shift from the baseline (unit ratio or zero log ratio) reflects copy number changes. Such changes can be identified in many ways, for example, one can use segmentation algorithms to partition the signal then try to classify such segments into gain, copy neutral and loss status.
Yet, for large datasets, one can take advantage of the signal distribution at each SNP, and cluster each individual from the distribution into a component that would reflect a given copy number change.
We developped a Gaussian Mixture Model, which detect copy number variation from the distribution of copy number ratios. From the data, it will fit one component for each of the following copy number states: deletion, copy-neutral, 1 and 2 additional copy; with a constraint on the difference between the mixture means. Then for a given individual, it will determine the probabilities for each copy number state and compute the expected copy number (dosage).
License
The GMM algorithm is licensed under the GNU General Public License, version 2 or later. For details, see http://www.gnu.org/licenses/old-licenses/gpl-2.0.html.
Usage
The GMM can be applied to identify CNVs from any rectangular matrix of copy number ratio.
Format is like : chr pos sample1 sample2 ...
Fields should be tab-delimited and it assumes data (within chromosome) are sorted by position.
An example input file is available within the GMM_CNV.zip (see Download section).
For Matlab users, download the source code and use the callCNVs.m script.
Users without Matlab, can use the compiled version and the Matlab Component Runtime (MCR). (Please note, we are only providing a compiled Linux x86_64 version for now).
Then you can use (and edit according to your need) the shell script called "run_CallCNVs.sh"
sh run_CallCNVs.sh path_to_the_MCR/v79/
Requirements
If you have the MATLAB software, you can directly use the source code.
Otherwise, you will need to download the Matlab Component Runtime to use the executables (see Download section).
Download
| Description | File Name | Size | md5sum |
|---|---|---|---|
| MCR for 64-bit Linux | MCR2007_x86_64.zip | 224M | 451c54a811b3e01402b6a46a1b814c4d |
| Linux Executables (+ example input file) | GMM_CNV.zip | 556k | bd579f39c340a50de2bb80a649643be3 |
| Source code | GMM_CNV_SOURCE.zip | 16k | 3cb7799bf3e180b33a6742ef382b105e |
| Example output files | GMM_CNV_outputs.zip | 460k | 6b621a6a8e279697f610db35810777ce |
Frequently Ask Questions
* What are the default component the model will try to fit?
The current implementation models deletion, copy neutral, 3 copies and more than 3 copies.
* What happen if the model fails to fit the data ?
You will model will move to the next SNP to process and you will simply get the warning :
Exiting: Maximum number of iterations has been exceeded - increase MaxIter option.
Missing data will be set as 0.
* I am getting :
Exiting: Maximum number of iterations has been exceeded - increase MaxIter option.
What does this mean? The model could not find the component separation before reaching its maximal iteration limit. This can be due to noisy data, or distribution where no such separation exists. Try increasing : MAX_FUN_CALL=10000; # nb of optimization function call MAX_FUN_ITER=5000; # nb of iterations for each optimization function call But note, this can significantly increase the runtime.
* Can I apply some extra normalization before fitting the Gaussian Mixture Model?
Yes, by default a Loess smoothing is applied. (This step can be skipped by setting DO_LOESS_SMOOTH=0 in the shell script or setting DO_LOESS=0; in callCNVs.m).
It is also recommended that adequate normalization is applied and that such normalized ratios are provided in the matrix input file.
* I am getting this error :
error while loading shared libraries: libmwmclmcrrt.so: cannot open shared object file: No such file or directory
what does it mean?
Most likely your LD_LIBRARY_PATH is not pointing correctly to the MCR. The run_[appName].sh script should do it for you.
sh run_callCNVs.sh /path-to-my-MCR/v79 test.dat # i.e. for compiled distrib with Matlab 2008, build v79
