Partitioned Heritability
In this tutorial, you will partition the heritability of BMI by functional category, using the baseline model. We will assume you have already followed the instructions in the README for downloading and installing python and ldsc.
The LD scores used in Finucane, Bulik-Sullivan et al., bioRxiv can be downloaded here. To do the basic baseline analysis, you will need to download the baseline model LD scores (baseline.* in 1000G_Phase1_baseline_ldscores.tgz), regression weights (weights.* in weights_hm3_no_hla.tgz), and allele frequencies (1000G.mac5eur.* in 1000G_Phase1_frq.tgz). To do a cell-type group comparison (described below) you will also have to download the cell type files (1000G_Phase1_cell_type_groups.tgz). Corresponding files built from 1000 Genomes Phase 3 are also available.
To download the BMI summary statistics, go to http://www.broadinstitute.org/collaboration/giant/index.php/GIANT_consortium_data_files and download GIANT_BMI_Speliotes2010_publicrelease_HapMapCeuFreq.txt. (You can use the download GZIP link and then unzip the resulting file.)
Check that the file head is:
> head GIANT_BMI_Speliotes2010_publicrelease_HapMapCeuFreq.txt
MarkerName Allele1 Allele2 Freq.Allele1.HapMapCEU p N
rs10 a c 0.0333 0.708 80566
rs1000000 g a 0.6333 0.506 123865
rs10000010 c t 0.425 0.736 123827
rs10000012 c g 0.8083 0.042 123809
rs10000013 c a 0.1667 0.0689 123863
rs10000017 t c 0.2333 0.457 123262
rs1000002 c t 0.475 0.0322 123783
rs10000023 t g 0.5917 0.939 123756
rs10000029 t c 0.975 0.24 103623
Next, convert this file to the .sumstats format (see the docs) using munge_sumstats.py. We recommend only keeping HapMap3 SNPs; to do this, you can download a list of HapMap3 SNPs here. Unzip this file to get w_hm3.snplist, and then run
python munge_sumstats.py --sumstats GIANT_BMI_Speliotes2010_publicrelease_HapMapCeuFreq.txt\
--merge-alleles w_hm3.snplist
--out BMI\
--a1-inc
This will give you a file called BMI.sumstats.gz.
The following command will allow you to partition heritability:
python ldsc.py
--h2 BMI.sumstats.gz\
--ref-ld-chr baseline.\
--w-ld-chr weights.\
--overlap-annot\
--frqfile-chr 1000G.mac5eur.\
--out BMI_baseline
Partitioning heritability with 53 overlapping categories takes about 10 minutes, mostly spent reading in all of the annotation matrices. Here is what each of the flags means:
This flag tells ldsc to compute partitioned heritability. The argument should be a file in the .sumstats format, which can be gzipped or not. This file should not include custom array data like immunochip, and it should be data from a population that is similar to the population in the LD scores. The LD scores downloaded for this tutorial are European LD scores.
The --ref-ld flag tells ldsc which LD Score files to use as the independent variable in the LD Score regression. The --ref-ld-chr flag is used for LD Score files split across chromosomes. By default, ldsc appends the chromosome number to the end. For example, typing --ref-ld-chr baseline. tells ldsc to use the files baseline.1.l2.ldscore.gz, ... , baseline.22.l2.ldscore.gz and baseline.1.l2.M_5_50, ... , baseline.22.l2.M_5_50. If the chromosome number is in the middle of the filename, you can tell ldsc where to insert the chromosome number by using an @ symbol. For example, --ref-ld-chr ld/chr@. The argument to --ref-ld should omit the file suffixes (.l2.ldscore.gz, .M_5_50).
If you are using the --overlap-annot flag, then you must have .annot.gz files with the same prefix, and these must be the .annot.gz files that were used as input to --l2 when computing the LD scores. In other words, your .annot.gz files must have a row for every SNP in your reference panel. On the other hand, your .l2.ldscore.gz files may have fewer rows if you used the --print-snps flag when they were computed.
You can input here either a single set of file or a comma-separated list of files. Below, there is an example with comma-separated files. When inputting multiple files, be extra careful not to introduce any co-linearity! For example, don't have the same annotation in two files, or a set of annotations in one file that forms a disjoint union of an annotation from the second file. This will cause ldsc to throw an error. This won't be a problem if you only use the files downloaded from the LDSCORE folder above. Also, if you are using your own annotation files, be sure that the union of all annotations you are including is the set of all variants, for example by including an annotation of all ones.
This flag gives the location of the LD scores used for regression weights. It should be a non-partitioned .l2.ldscore.gz file that was computed using the regression SNPs. We recommend using HapMap3 SNPs, excluding the HLA region, as default regression SNPs; that is what is in weights.*.
This flag tells ldsc where to print the the output. It will append .log and .results to this prefix. So in this case, you will see BMI_baseline.log and BMI_baseline.results.
This flag tells ldsc that the categories you used to generate baseline.*.l2.ldscore.gz overlap with each other. To interpret the results when there are overlapping categories, ldsc needs to read in the .annot.gz files. It does this using the same prefix as in --ref-ld or --ref-ld-chr.
In order to use only SNPs with MAF > 5% (see discussion of 22.M_5_50), ldsc needs to know how many SNPs with MAF > 5% there are in every pairwise intersection of categories. Because the same annotations may be used in different combinations--for example, several different reference panels may be input at the same time--a single .M_5_50 file doesn't suffix. Thus --overlap-annot requires a frequency file, rather than just a .M_5_50 file. If you use the --not-M-5-50 flag, then this file is not necessary.
Format for overlapping categories changing soon.
This file has the results of the analysis in tab-delimited form. If any category contains all SNPs, then that category will not appear in this file. There is one row for each category and columns summarizing the results: Proportion of SNPs, Proportion of heritability, Enrichment, and standard errors. Enrichment is (Prop. heritability) / (Prop. SNPs). If you use the --print-coefficients flag, then there will also be columns for the regression coefficients. (See Finucane, Bulik-Sullivan et al., bioRxiv for a discussion of the relationship between the coefficients and proportions of heritability.)
Update: a more efficient way to do cell type-specific analyses is now implemented and described in the Cell type-specific analyses section. We only recommend the approach described here if you are interested not just in prioritizing relevant cell types but also in estimating proportion of heritability.
==
To compare cell-type groups, we use the coefficient z-score in an analysis containing the full baseline model. This means that we are controlling for the 53 categories of the full baseline model when comparing cell-type groups. To run the analysis for a single cell-type group, say CNS, run:
python ldsc.py
--h2 BMI.sumstats.gz\
--w-ld-chr weights.\
--ref-ld-chr CNS.,baseline.\
--overlap-annot\
--frqfile-chr 1000G.mac5eur.\
--out BMI_CNS\
--print-coefficients
Note that here, we are using a comma-separated list of file prefixes with --ref-ld-chr, and we also include the --print-coefficients flag so that we can compare coefficients as well as proportions of heritability. In Finucane, Bulik-Sullivan et al., bioRxiv, we rank cell types using the z-score of the coefficient of the cell type.