This repository contains the python implementation of networkGWAS method. The paper underlying this method is available at https://doi.org/10.1093/bioinformatics/btad370. networkGWAS main steps are showed in the following figure:
Each of the different steps of networkGWAS can be computed by the following scripts:
- neighborhood aggregation of the SNPs according to the biological network (1_nb_aggregation.py);
- 2-level permutation procedure, which combines a circular permutation of the SNPs and degree-preserving permutation of the network (2_circPerm_nwPerm.py);
- calculation of the log-likelihood ratio (lrt) test statistics on both the non-permuted setting and the permuted settings by employing the modified FaST-LMM snp-set function, which is available in the folder "LMM" (3_LMM.py);
- calculation of the p-values. Per each neighborhood, this is done by: (i) obtaining the null distribution through the pooling of the lrt statitistics on the permuted settings, (ii) calculating the ratio of statistics in the null distribution that are larger than or equal to the statistic for the neighborhood obtained on the non-permuted setting (4_obtain_pvals.py).
- identifying the statistically associated neighborhood by means of the Benjamini-Hochberg (B-H) procedure in case of analysing one phenotype only or by using the hierarchical procedure based on B-H procedure in case of multiple phenotypes (5_associated_neighborhoods.py).
Available here a toy-dataset on which try the method. Detail below.
The toy dataset is comprised of:
- genotype file in Plink .bed/.bim/.fam format ("data/genotype.bed", "data/genotype.bim", "data/genotype.fam")
- simulated phenotype in Plink .pheno format ("data/y_50.pheno")
- the adjacency matrix of the biological network, i.e., the protein-protein interaction (PPI) network in Pickle format ("data/PPI_adj.pkl"). In particular, the adjacency matrix is saved as a pandas dataframe, where the index and columns are the names of the genes, which represents the nodes of the PPI network.
- the dictionary presenting the nodes (gene) as keys with as values a numpy boolean vector that presents True where the SNPs (ordered according to the .bim file) are mapped onto that particular gene ("data/gene_snps_index.pkl")
For more details on the Plink formats, please refer to https://www.cog-genomics.org/plink/2.0/formats.
In case the user has a PPI network in two columns format (without header), that is:
gene1 gene4
gene1 gene2
... ...
it is possible to convert it to the required dataframe (i.e., adjacency matrix) format using the command:
python3 convert_ppi.py --i data/ppi_edges.txt --o data/PPI_adj.pkl
The code only supports python3 and requires the following packages and submodules:
- numpy (tested on 1.18.1)
- pandas (tested on 1.0.1)
- fastlmm (https://github.com/fastlmm/FaST-LMM) (tested on 0.4.8)
- pysnotools (tested on 0.4.19)
python3 1_nb_aggregation.py \
--i data \
--o results/settings \
--g2s gene_snps_index.pkl \
--bim genotype.bim \
--nw PPI_adj.pkl \
--nbs neighborhoods.txt
python3 2_circPerm_nwPerm.py \
--i data \
--g2s gene_snps_index.pkl \
--bim genotype.bim \
--nw PPI_adj.pkl \
--perm 100 \
--alpha 0.5 \
--seed 42 \
--onwdir results/settings/permutations/networks/ \
--onbdir results/settings/permutations/neighborhoods/ \
--onw nw_ \
--onb nbs_
Since this command runs the generation of 100 permuted settings, it might take a few minutes, depending on the resources available.
python3 3_LMM.py \
--genotype data/genotype \
--phenotype data/y_50.pheno \
--nbs results/settings/neighborhoods.txt \
--kernel lin \
--odir results/llr/ \
--ofile llr.pkl
3.2) running the lrt calculation on one of the permuted settings, i.e., the permuted setting with index (or permutation id) j = 0:
python3 3_LMM.py \
--genotype data/genotype \
--phenotype data/y_50.pheno \
--nbs results/settings/permutations/neighborhoods/nbs_ \
--kernel lin \
--j 0 \
--odir results/llr/permuted \
--ofile llr_
IMPORTANT: This command has to be run for enough permutations, e.g., in this example we use 100 permutations. This can be done by changing the value of the input parameter --j with a value from 0 to the maximum permutation id available, which in this example would be 99. Note that since the permutation id (--j) is a command line argument, 3_LMM.py for the different permuted settings can be run in parallel. When enough resources should not be available to run multiple permutations in parallel, it is possible to run the lrt calculation sequentially by specifying a range of permutations. The following command calculates the lrt values on the permuted settings from 0 to 99:
python3 3_LMM.py \
--genotype data/genotype \
--phenotype data/y_50.pheno \
--nbs results/settings/permutations/neighborhoods/nbs_ \
--kernel lin \
--j 0 99 \
--odir results/llr/permuted \
--ofile llr_
python3 4_obtain_pvals.py \
--inpath results/llr/llr.pkl \
--inpathperm results/llr/permuted/llr_ \
--nperm 100 \
--dirnd results/null_distr/ \
--dirpv results/pvals/ \
--fignd null_distr.png \
--figpv qqplot.png \
--outpathnd null_distr.pkl \
--outpathpv pvals.pkl
python3 5_associated_neighborhoods.py \
--nw 'data/PPI_adj.pkl' \
--pv 'results/pvals/pvals.pkl' \
--q1 0.05
or identifying the statistically associated neighborhoods in case of multiple phenotypes that are from the same study:
python3 5_associated_neighborhoods.py \
--nw 'data/PPI_adj.pkl' \
--pv 'results/pvals/pvals_pheno1.pkl' 'results/pvals/pvals_pheno2.pkl' 'results/pvals/pvals_pheno3.pkl' \
--q1 0.05 \
--q2 0.05
Note that this last command would work only after having obtained the p-values for "pheno1'', "pheno2'', and "pheno3''.
It is also possible to run the entire analysis using one script, namely:
bash run_analysis.sh
The user would need to change the input files and parameters in the file according to their data.
In order to reproduce the results presented in the manuscript, here a list of the data availabilities:
- genotype: the full imputed version of the A. thaliana genotype is available on the AraGWAS database (https://aragwas.1001genomes.org/#/download-center); genetic variants presenting MAF < 0.05 have been filtered out;
- phenotype: simulated according to the procedure detailed in [1];
- PPI network: the PPI network is downloaded from the The Arabidopsis Information Resource (TAIR) database (https://www.arabidopsis.org/).
- genotype: simulated using sim1000G package [4] giving as input the VCF from Phase III 1000 genomes sequencing data; the simulated SNPs present MAF < 0.1;
- phenotype: simulated according to the procedure detailed in [1];
- PPI network: the PPI network is downloaded from the STRING database (https://string-db.org/) by selecting the high confidence PPIs (score >= 700) for the model organism H. sapiens.
- genotype: the full imputed version of the A. thaliana genotype is available on the AraGWAS database (https://aragwas.1001genomes.org/#/download-center); genetic variants presenting MAF < 0.05 have been filtered out;
- phenotype: the natural phenotypes for A. thaliana have been downloaded from the AraPheno database (https://arapheno.1001genomes.org/phenotypes/);
- PPI network: the PPI network is downloaded from the STRING database (https://string-db.org/) by selecting the high confidence PPIs (score >= 700) for the model organism A. thaliana.
- genotype & phenotype: genotype and phenotype for the model organism S. cerevisiae are available at http://1002genomes.u-strasbg.fr/files/, which is the repository of the publication [3];
- PPI network: the PPI network is downloaded from the STRING database (https://string-db.org/) by selecting the high confidence PPIs (score >= 700) for the model organism S. cerevisiae.
- genotype & phenotype: genotype and phenotype are obtained from the Estonian BioBank dataset (https://genomics.ut.ee/en/content/estonian-biobank);
- PPI network: the PPI network is downloaded from the STRING database (https://string-db.org/) by selecting the high confidence PPIs (score >= 700) for H. sapiens.
[1] G. Muzio, L. O’Bray, L. Meng-Papaxanthos, J. Klatt, K. Fischer, K. Borgwardt, networkGWAS: a network-based approach to discover genetic associations, Bioinformatics 9(6), 2023.
[2] C. Lippert, J. Xiang, D. Horta, C. Widmer, C. Kadie, D. Heckerman, J. Listgarten, Greater power and computational efficiency for kernel-based association testing of sets of genetic variants, Bioinformatics 30(22):3206–3214, 2014.
[3] J. Peter, M. De Chiara, A. Friedrich et al., Genome evolution across 1,011 Saccharomyces cerevisiae isolates. Nature 556, 339–344, 2018.
[4] A. Dimitromanolakis, J. Xu, A. Krol, L. Briollais, sim1000g: a user-friendly genetic variant simulator in R for unrelated individuals and family-based designs. BMC Bioinformatics 20(26), 2019.
This project has received funding from the European Union’s Horizon 2020 research and innovation programme under the Marie Skłodowska-Curie grant agreement No 813533. This study was supported in part by the Alfried Krupp Prize for Young University Teachers of the Alfried Krupp von Bohlen und Halbach-Stiftung (Borgwardt).