This repository holds the code for the paper "PALM: A powerful and adaptive latent model for prioritizing risk variants with functional annotations".
Many complex phenotypes are highly polygenic, making a large number of associated genetic variants fail to be discovered in GWAS. Evidence shows that associated variants are enriched in some functional annotations, which offers us an opportunity to increase the statistical power of large-scale multiple testing by properly leveraging these auxiliary information. Existing methods for integrating functional annotations with GWAS results have limitations: (1) most of the existing methods assume a linear model for risk variant identification; (2) few of existing methods are scalable to handle a large number of annotations while maintaining good interpretability.
In this work, we proposed a novel statistical method to better utilize functional annotations in risk variants prioritization. Specifially, we assume the p-values from GWAS summary statistics follow a two-groups model with SNP-specific priors; we model the SNP-specific priors in the logit scale using boosted trees. An EM algorithm was developed to simultaneously optimize model parametes and add trees in a stagewise fashion such that annotations are adaptively incorporated into the model. The merits of boosted trees transplant to PALM: not only does it allow complex relationship between functional annotations and SNP association status but also automatically selects important annotations which provides some biological insights. Simulation studies and real data analysis illustrated PALM's superior performance on risk variants prioritization compared with existing representative methods and its effectiveness in identifying enriched cell-type/tissue specific annotations.
- Git clone the repository and install the package.
$ git clone https://github.com/YangLabHKUST/PALM.git
$ cd PALM
$ conda env create -f environment.yml
$ conda activate palm- Check the installation status:
$ python PALM/prioritize.py -h- Test if PALM runs successfully:
$ python test.pyInput files of PALM include:
- GWAS summary statistics file
- Annotation file
The GWAS summary statistics file should contain a SNP column and a pvalue column, e.g,
$ head Bipolar_Disorder.csv
SNP CHR BP MAF A1 A2 pvalue
rs10907175 1 1130727 0.084493 A C 0.351999999999945
rs2887286 1 1156131 0.175944 T C 0.390399999999947
rs6685064 1 1211292 0.0666004 C T 0.943799999999981
rs1571150 1 1474304 0.333002 C A 0.667700000000157
rs7290 1 1477244 0.306163 T C 0.847799999999996
rs3766180 1 1478153 0.306163 T C 0.852900000000085
rs3766178 1 1478180 0.306163 T C 0.801899999999967
rs7533 1 1479333 0.306163 A G 0.85810000000001
rs7517401 1 1483010 0.349901 G A 0.700699999999828The annotation file has a SNP column for merging with the GWAS file and the other columns are different annotations, e.g,
$ head region9.csv
downstream exonic intergenic intronic ncRNA_exonic ncRNA_intronic upstream UTR3 UTR5 SNP
0 0 1 0 0 0 0 0 0 rs1000000
0 0 0 1 0 0 0 0 0 rs10000010
0 0 0 1 0 0 0 0 0 rs10000023
0 0 1 0 0 0 0 0 0 rs1000003
0 0 1 0 0 0 0 0 0 rs10000033
0 0 0 1 0 0 0 0 0 rs10000037
0 0 0 0 0 1 0 0 0 rs10000041
0 0 1 0 0 0 0 0 0 rs1000007
0 0 1 0 0 0 0 0 0 rs10000075PALM will integrate the GWAS summary statistics with annotations to prioritize risk SNPs. Specify <data_dir> containing the input files and <out_dir> for saving output files. The default method in the prior model is boosted trees with tree depth = 2 and shrinkage parameter = 0.1. By default, PALM will perform 5-fold cross-validation with maximum number of iterations 3000 to select the optimal number of iterations for the final model.
python ./PALM/prioritize.py \
--out <out_dir>/Bipolar_Disorder \
--sumstats <data_dir>/Bipolar_Disorder.csv \
--annotation <data_dir>/region9.csv \
--model tree --depth 2 --eta 0.1 \
--max-iter 3000 --nfold 5 -vThe main script prioritize.py will add two columns post and assoc to the input GWAS file and save as the new dataframe as the prioritization result. post represents the posterior probability of one SNP being associated with the phenotype given its p-value and annotations; assoc represents the SNP association status under a certain FDR control level (by default, 0.1). Besides, a file recording the log-likelihoods on test folds and a file recording computational times will be saved. And if specified, the estimated α in the two-groups model and the evaluated variable importance will be saved as well.
An example of the major output file:
$ head Bipolar_Disorder-J2-v0.1-PALM-tree-prioritize.csv
SNP CHR BP MAF A1 A2 pvalue post assoc
rs10907175 1 1130727 0.084493 A C 0.351999999999945 0.2459430394857876 0.0
rs2887286 1 1156131 0.175944 T C 0.390399999999947 0.21100387604316348 0.0
rs6685064 1 1211292 0.0666004 C T 0.943799999999981 0.10466973830102834 0.0
rs1571150 1 1474304 0.333002 C A 0.667700000000157 0.312529201925508 0.0
rs7290 1 1477244 0.306163 T C 0.847799999999996 0.22741445464179644 0.0
rs3766180 1 1478153 0.306163 T C 0.852900000000085 0.2170770906036481 0.0
rs3766178 1 1478180 0.306163 T C 0.801899999999967 0.2209031523383583 0.0
rs7533 1 1479333 0.306163 A G 0.85810000000001 0.1511410932793709 0.0
rs7517401 1 1483010 0.349901 G A 0.700699999999828 0.21465550994232505 0.0We provide the source codes for reproducing the experimental results of PALM. The 30 GWAS summary statistics and functional annotations in real data analysis can be downloaded here.
- Simulation script and simulation results
- Supplemental simulation scripts and supplemental simulation results
- Real data format and an example on real data
- Real data analysis script and real data analysis results
The prior model of PALM is a general framework which allows us to apply any appropriate method to fit the logit-scale prior probabilities of the SNP association status. For example, the boosted trees can be replaced with a neural network. We also provide the implementation of network-based model with a 3-layer fully connected network. A demo on simulated data with network-based model can be found here.
After conducting some simulations, we found that the performance of network-based model was sensitive to the network design and hyperparameters tuning (at least for the very simple multi-layer perceptron). Since network-based model is not robust in FDR control, we haven't applied it to real data. For risk variant prioritization, currently we suggest tree-based models. Further research is needed for the potential usage of neural network in genomic integrative analysis.
Xinyi Yu, Jiashun Xiao, Mingxuan Cai, Yuling Jiao, Xiang Wan, Jin Liu, Can Yang. PALM: A Powerful and Adaptive Latent Model for Prioritizing Risk Variants with Functional Annotations Bioinformatics. 2023. doi: https://doi.org/10.1093/bioinformatics/btad068.
Please contact Xinyi Yu ([email protected]) and Prof. Can Yang ([email protected]) if any enquiry.