PyFM

PyFM uses Bayes Factor-based modeling of different configurations of given variants for fine mapping. The framework is essentialy based on CaviarBF [1]. But like FINEMAP [2] our method further uses Shotgun Stochastic Search to explore the configurations-space more intelligently.

We copied the example data from CaviarBF and FineMap into the example subdirectory.

We prepared 2 example runs from command-line. Example 1's PIP and Rho match up nearly exactly with the CaviarBF results (difference arise from the Cholesky Decomposition round-offs). Example 2's Rho values don't match up exactly, but they are very close and the difference is likely due to different Cholesky Decomposition packages.

Example 1

python src/fm.py \
  -z example/eQTL/region.LOC284581.chr1.205831207.205865215.dosage.p1e-12.z \
  -r example/eQTL/region.LOC284581.chr1.205831207.205865215.dosage.p1e-12.ld \
  -o pyfm_results \
  -n 471 -c 2 -t 0 -e 0.1 -a 1.6

Example 1 using Shotgun Stochastic Search

python src/fm.py \
  -z example/eQTL/region.LOC284581.chr1.205831207.205865215.dosage.p1e-12.z \
  -r example/eQTL/region.LOC284581.chr1.205831207.205865215.dosage.p1e-12.ld \
  -o pyfm_results \
  --configs-method SSSConfigurations \
  --SSS-iterations 100 \
  -n 471 -c 2 -t 0 -e 0.1 -a 1.6

Example 1 with up to 5 Causal Variants

python src/fm.py \
  -z example/eQTL/region.LOC284581.chr1.205831207.205865215.dosage.p1e-12.z \
  -r example/eQTL/region.LOC284581.chr1.205831207.205865215.dosage.p1e-12.ld \
  -o pyfm_results \
  --configs-method SSSConfigurations \
  --SSS-iterations 100 \
  -n 471 -c 5 -t 0 -e 0.1 -a 1.6

Example 2

python src/fm.py \
  -z example/finemap_examples/finemap_data.z \
  -r example/finemap_examples/finemap_data.ld \
  -o pyfm_results \
  -n 5363 -c 2 -t 0 -e 0.1 -a 1.6

Arguments:

For further explanation of the arguments, check CaviarBF arguments since they are very similar: CaviarBF Manual (if does not open on webpage, please download the PDF file from above and open locally)

Argument	Type	Description
-z	FILE	zfile
-r	FILE	rfile containing pairwise LD correlation matrix
-o	DIR	output directory
-n	int	(default: 0) sample number
-c	int	(default: 3) maximum number of causal variants considered
-t	int	(default: 0) prior type
-e	float	(default: 0) epsilon, noise factor added to correlation matrix
-a	[float]	(default: 0.1 0.2 0.4 0.8 1.6) priors to run on
-p	float	(default: 1; i.e. reports all) rho cutoff to be used
--exclude_null	bool	(optional) if used, null model score
--configs-method	{AllConfigurations,SSSConfigurations}	(default: AllConfigurations) causal configurations exploration method
--SSS-iterations	int	(default: 100) Number of iterations for Shotgun Stochastic Search to run
--SSS-alpha1	float	(default: 1.5) Temperature Parameter for SSS sampling stage 1. When this parameter is high, the chosen model in each group will be the highest scoring model (greedy search). If this parameter is low, the chosen model in each group will be more random.
--SSS-alpha2	float	(default: 1.5) Temperature Parameter for SSS sampling stage 2. When this parameter is high, the group chosen will be the highest scoring model (greedy search). If this parameter is low, the group chosen will be more random.
--Random-Seed	float	(default: 'None') Random Seed For Shotgun Stochastic Search. Set to 'None' For No Random Seeding

Compare results with CaviarBF

Download CaviarBF from by

git clone https://bitbucket.org/Wenan/caviarbf.git

Install by running the Makefile using

make

Run CaviarBF on the example files. Note: CaviarBF has two modules, caviarbf and model_search

caviarbf builds the Bayes factors for each SNP, and model_search find the best model of SNP combinations based on exhaustive/greedy search. Again, please refer to CaviarBF Manual.

We had prepared the example code for the following file structures

caviarbf/
    src/
    Makefile
    caviarbf (executable)
    ...
PyFM/
    src/
        fm.py (executable)
        ...
    example/
    pyfm_results/
    caviarbf_results/
    ...

CaviarBF: caviarbf module

Don't know why sometimes it returns killed: 9, but just run it again.

Similar to PyFM, but -o is a PATH to the FILE, instead of PATH to DIR

Example 1a

../caviarbf/caviarbf \
  -z example/eQTL/region.LOC284581.chr1.205831207.205865215.dosage.p1e-12.z \
  -r example/eQTL/region.LOC284581.chr1.205831207.205865215.dosage.p1e-12.ld \
  -o caviarbf_results/region.LOC284581.chr1.205831207.205865215.dosage.p1e-12.bf \
  -n 471 -c 2 -t 0 -e 0.1 -a 1.6

Example 2a

../caviarbf/caviarbf \
  -z example/finemap_examples/finemap_data.z \
  -r example/finemap_examples/finemap_data.ld \
  -o caviarbf_results/finemap_example.bf \
  -n 5363 -c 2 -t 0 -e 0.1 -a 1.6

CaviarBF: build_model module

-e for exhaustive search, and -s for greedy stepwise search

Example 1b

../caviarbf/model_search \
  -i caviarbf_results/region.LOC284581.chr1.205831207.205865215.dosage.p1e-12.bf \
  -o caviarbf_results/test_stepwise \
  -s -m 237 -p 0 > caviarbf_results/log.txt

Example 2b

../caviarbf/model_search \
  -i caviarbf_results/finemap_example.bf \
  -o caviarbf_results/test_finemap_stepwise \
  -s -m 55 -p 0 > caviarbf_results/log.txt

FINEMAP

FINEMAP can be downloaded and ran directly. This run shows one of its compiled example.

./finemap_v1.4.2_x86_64 \
  --sss \
  --in-files example/data \
  --dataset 1

Benchmark against FINEMAP

Batch code for simulation and running the different tools are in the simulation folder. Simulation was done using Elison, Weston's tool. Analyses are in the generate_plots notebook.

Known Issues

1. Potential memory leak when running k >= 5 in SSS mode
2. Certain ld-file and simulated z-file can cause issue during Cholesky Decomposition, result in halting SSS mode. Simply re-simulate those cases with slightly different noise.
3. Longer runtime than CaviarBF and FINEMAP, in exhaustive and SSS mode, respectively, due to lack of optimization and coded in Python

References

• [1] Chen W, Larrabee BR, Ovsyannikova IG, Kennedy RB, Haralambieva IH, Poland GA, Schaid DJ. Fine Mapping Causal Variants with an Approximate Bayesian Method Using Marginal Test Statistics. Genetics. 2015 Jul;200(3):719-36. doi: 10.1534/genetics.115.176107. Epub 2015 May 6. PMID: 25948564; PMCID: PMC4512539.
• [2] Christian Benner, Chris C.A. Spencer, Aki S. Havulinna, Veikko Salomaa, Samuli Ripatti, Matti Pirinen, FINEMAP: efficient variable selection using summary data from genome-wide association studies, Bioinformatics, Volume 32, Issue 10, May 2016, Pages 1493–1501, https://doi.org/10.1093/bioinformatics/btw018
• [3] Elison, Weston, CSE_284_Finemapping (2024), GitHub repository, https://github.com/westonelison/CSE_284_Finemapping
• [4] Fortune MD, Wallace C. simGWAS: a fast method for simulation of large scale case-control GWAS summary statistics. Bioinformatics. 2019 Jun 1;35(11):1901-1906. doi: 10.1093/bioinformatics/bty898. PMID: 30371734; PMCID: PMC6546134.