🌓 stratAS

A method for allele specific analyses across cell states and conditions. stratAS implements population-scale tests and predictive models for allele-specific activity (typically RNA or ATAC data).

Workflow

Input data / dependencies

• A .vcf file containing phased individual genotypes and an "AS" field listing the allelic counts. This gets turned into a flat file to be loaded in the --input flag (see below).
• A samples file with the columns ID and CONDITION that is in the same order as the vcf input. CONDITION is currently restricted to 0/1. This file goes into the --samples flag.
• A .bed file listing the physical positions of the features to be tested for allelic imbalance. This file goes into the --peaks flag.
• Global/local parameter files generated by params.R (see below). At minimum a global overdispersion parameter file with the columns ID and PHI with entries in any order goes into the --global_params flag.
• R, and the VGAM, optparse libraries installed.

Recommended pre-processing and allelic counts

• Phase and impute your genotype data with EAGLE on the Haplotype Reference Consortium imputation server.
• Process all sequence data with the WASP mapping pipeline.
• Use the GATK ASEReadCounter to count reads then convert into vcf (see pipeline/ scripts).

Estimating prior parameters

params.R infers the read count distribution (beta binomial) parameters and needs to be run on a whole genome vcf one individual at a time. The input is an --inp_counts file which must contain the headers CHR POS HAP REF.READS ALT.READS where HAP is the 0|1 or 1|0 vcf haplotype code.

A vcf file containing one individual is converted to counts as follows:

zcat $VCF \
| grep -v '#' \
| cut -f 1,2,10 \
| tr ':' '\t' \
| awk '$3 != "1|1" && $3 != "0|0"' \
| cut -f 1-3,7 \
| tr ',' '\t' \
| sed 's/chr//' \
| awk 'BEGIN { print "CHR POS HAP REF.READS ALT.READS" } $4 + $5 > 0' \
 > $OUT.counts

For tumor data, local CNV-specific parameters are also estimated, and the --inp_cnv file must be provided with headers CHR P0 P1 listing the boundaries of CNV regions. This file can additionally contain a CNV column listing the CNV estimate (centered to zero as in TCGA calls) for inclusion as a covariate in the final analysis.

inference is then performed by running:

Rscript params.R --min_cov 5 --inp_counts $OUT.counts --inp_cnv $CNV --out $OUT

A $OUT.global.params file is generated containing the parameters (with header PHI MU N) and (optionally) an $OUT.local.params file is generated containing the positions and parameters for each CNV (with header CHR P0 P1 PHI MU N).

Testing for population allele specificity

stratas.R computes the AS statistics from a VCF of all individuals and the prior parameters estimated above.

A vcf file containing all individuals is converted to counts and split into batches as follows:

zcat $VCF\
| grep -v '#' \
| cut -f 1-5,9- \
| tr ':' '\t' \
| awk '{ for(i=1;i<=5;i++) { printf "%s ",$i; } for(i=6;i<=10;i++) { if($i == "GT") gtnum=i-6; if($i=="AS") asnum=i-6; } for(i=11;i<=NF;i++) { if( (i-11) % 5 == asnum || (i-11) % 5 == gtnum ) printf " %s",$i; } print ""; }' \
| sed 's/[|,]/ /g' | split -d -l 15000 - $OUT.MAT.split.

Each batch is then processed as follows:

Rscript stratas.R \
--input $OUT.MAT.split.00 \
--samples SAMPLES.ID \
--peaks gencode.protein_coding.transcripts.bed \
--global_param GLOBAL.params \
--local_param LOCAL.params \

Learning predictive models for TWAS/RWAS/CWAS analysis

stratas.R can compute multi-variate predictive models for downstream integration with GWAS data, using both allelic and total activity.

Inputs are similar to the tests for allele specificity described above and additionally require a total activity matrix, covariates, and the predict flags:

Rscript stratas.R \
--input $OUT.MAT.split.00 \
--samples SAMPLES.ID \
--peaks gencode.protein_coding.transcripts.bed \
--global_param GLOBAL.params \
--local_param LOCAL.params \
--total_matrix KIRC.gexp \
--covar KIRC.gexp.cov \
--predict_snps HM3.extract \
--predict --predict_only > $OUT.profile

This generates predictive model files for every feature in the TWAS/FUSION format, which can then be integrated with GWAS data using the FUSION software. The generated model types inside each model file are "lasso", "lasso.as", "lasso.combined", "top1.as", "top1", "top1.combined"; where lasso versus top corresponds to full locus penalized versus top SNP models; and *, *.as, *.combined indicates models using total activity, allele-specific activity, or both jointly. An $OUT.profile file is generated which lists model performance characteristics.

Important: This analysis differs from the individual allele-specific tests in two ways: (1) the CONDITION value is not used and one model is trained on all samples; (2) models are still trained even if no allele-specific informatino is available. To restrict to models with both sources of signal (which are likely to be more accurate as a group), run the allele-specific tests described above and retain only peaks that have non-NA values.

Example total activity input files (KIRC.gexp, KIRC.gexp.cov, and HM3.extract) based on TCGA data are provided in the example/ directory.

Output data

By default, the outputs are printed to screen with each line containing the following entries:

Column	Description
CHR	Chromosome
POS	Position of test SNP
RSID	ID of test SNP
P0	Start of gene/peak
P1	End of gene/peak
NAME	Name of gene/peak
CENTER	Center position of peak (or TSS for gene)
N.HET	# of heterozygous individuals tested
N.READS	# of reads tested in total
ALL.AF	Allelic fraction estimate from beta binomial test across both conditions
ALL.BBINOM.P	Beta-binomial test for imbalance across both conditions
C0.AF	Allelic fraction estimate from condition 0
C0.BBINOM.P	Beta-binomial test for imbalance in condition 0
C1.AF	Allelic fraction estimate from condition 1
C1.BBINOM.P	Beta-binomial test for imbalance in condition 1
DIFF.BBINOM.P	Beta-binomial test for difference between conditions

Enabling the --total_matrix flag additionally runs a standard eQTL test (and interaction term), producing the following columns:

Column	Description
ALL.TOTAL.Z	Test statisic from linear model across all samples
ALL.TOTAL.P	P-value from linear model across all samples
C0.TOTAL.Z	Test statisic from linear model across condition 0
C0.TOTAL.P	P-value from linear model across condition 0
C1.TOTAL.Z	Test statisic from linear model across condition 1
C1.TOTAL.P	P-value from linear model across condition 1
DIFF.TOTAL.Z	Test statistic for difference/interaction between conditions
DIFF.TOTAL.P	P-value for difference/interaction between conditions

Enabling the --combine flag additionally estimates the (inverse-variance weighted) combined statistics from the allele-specific and total models, producing the following columns:

Column	Description
ALL.COMBINE.Z	Test statisic from linear model across all samples
ALL.COMBINE.P	P-value from linear model across all samples
C0.COMBINE.Z	Test statisic from linear model across condition 0
C0.COMBINE.P	P-value from linear model across condition 0
C1.COMBINE.Z	Test statisic from linear model across condition 1
C1.COMBINE.P	P-value from linear model across condition 1
DIFF.COMBINE.Z	Test statistic for difference between conditions
DIFF.COMBINE.P	P-value for difference between conditions

Enabling the --binom flag additionally runs a standard binomial test (and Fisher's test for interaction), producing the following columns:

Column	Description
ALL.BINOM.P	Binomial test for imbalance across both conditions
ALL.C0.BINOM.P	Binomial test for imbalance in condition 0
ALL.C1.BINOM.P	Binomial test for imbalance in condition 1
FISHER.OR	Fisher's test odd's ratio for difference between conditions
FISHER.DIFF.P	Fisher's test difference between conditions

Enabling the --bbreg flag additionally runs a beta binomial regression with CNV as covariate, and produces the following columns:

Column	Description
ALL.BBREG.P	Beta binomial regression (with covariates) for imbalance across both conditions
DIFF.BBREG.P	Beta binomial regression (with covariates) for imbalance difference between conditions
CNV.BBREG.P	Beta binomial regression (with covariates) for imbalance along CNV covariate

Enabling the --indiv flag additionally produces the following columns:

Column	Description
IND.C0	Number of each condition 0 individual included in this test (comma separated)
IND.C0.COUNT.REF	condition 0 REF allele counts of each individual included in this test (comma separated)
IND.C0.COUNT.ALT	condition 0 ALT allele counts of each individual included in this test (comma separated)
IND.C1	Number of each condition 1 individual included in this test (comma separated)
IND.C1.COUNT.REF	condition 1 REF allele counts of each individual included in this test (comma separated)
IND.C1.COUNT.ALT	condition 1 ALT allele counts of each individual included in this test (comma separated)

Example

An example locus with significant AS associations can be run by calling:

Rscript stratas.R \
--input example/ENSG00000075240.12.mat \
--samples example/KIRC.ALL.AS.PHE \
--peaks example/ENSG00000075240.12.bed \
--global_param example/KIRC.ALL.AS.CNV \
--local_param=example/KIRC.ALL.AS.CNVLOCAL

Detailed Parameters

All parameters can be displayed by running --help.

Basic input/output

Parameter
--input	Path to input file
--samples	Path to sample identifier file, must have ID and CONDITION columns
--peaks	Path to file containing peak/gene boundaries, must contain CHR P0 P1 NAME CENTER columns
--global_param	Path to global overdispersion parameter file
--local_param	Path to local overdispersion parameter file
--out	Path to output

Inclusion/exclusion

Parameter
--exclude	The mimium distance between SNPs allowed in the haplotype (to exclude variants in the same read)
--keep	Path to file listing samples to retain for analyses
--min_cov	Individuals must have at least this many reads (for both alleles) to be tested
--min_het	Minimum minor heterozygous frequency for test SNP
--min_maf	Minimum minor allele frequency for test SNP
--max_rho	Maximum local/global over-dispersion parameter for which to include individual in test
--window	Window (in bp) for SNPs to test around the peak boundary

Total (non-allelic) input

Parameter
--covar	Path to covariates for total activity
--total_matrix	Path to matrix of total activity, enables the linear model. Note: if --local_param is on, then CNV is included as covariate
--total_rn	Rank normalize the total expression phenotype

Prediction

Parameter
--predict	Build predictive models for each feature (for downstream TWAS/RWAS/CWAS analysis)
--predict_only	Do not perform individual allele-specific tests
--predict_snps	Path to file listing SNPs for which predictor weights should be fit (typically common HapMap3 SNPs)

Testing

Parameter
--bbreg	Also perform a beta binomial regression with local CNV status as a covariate (must also provide --local_param file)
--binom	Also perform a standard binomial test
--collapse_reads	Merge all sites for each individual prior to analysis (not recommended)
--combine	Output combined BBINOM and TOTAL statistics by Stouffer's method, must have all relevant flags for basic and total input
--mbased	Also perform the MBASED test for differences (requires MBASED libraries)

Copy number

Parameter
--cond_cnv_mean	Use MU as the covariate for BBREG, otherwise uses CNV as the covariate (requires --bbreg)
--fill_cnv	Set individuals with missing CNV calls to diploid and rho=0.01 (assuming they are neutral)
--mask_cnv_cutoff	Mask out any sites that have an absolute CNV value above this cutoff (NA = no masking)
--min_cnv_round	Absolute CNV values below this cutoff get set to zero

Simulation

Parameter
--sim	Simulate imbalance based on the input data and test it. Allelic-fraction specified by --sim_af
--sim_af1	Specify the allelic fraction for CONDITION==1
--sim_af0	Specify the allelic fraction for CONDITION==0
--sim_cnv	Add local CNVs to simulations
--sim_cnv_allelic	CNVs always impact one allele

Extras

Parameter
--indiv	Also report the per-individual allele fractions (Warning, this can produce large files, typically used for visualization of one feature)
--perm	# of permutations to shuffle the allele labels (0=off)
--perm_cond	# of permutations to shuffle the condition labels (0=off)
--seed	Random seed (for simulations/permtuations)