PCAngsd

PCAngsd is a program that estimates the covariance matrix for low depth next-generation sequencing (NGS) data in structured/heterogeneous populations using principal component analysis (PCA) to perform multiple population genetic analyses using an iterative procedure based on genotype likelihoods.

Since version 0.98, PCAngsd was re-written to be based on Cython for computational bottlenecks and parallelization and is now compatible with any newer Python version.

The method was published in 2018 and can be found here: [1]

Simulated low depth NGS data of 3 populations

Overview

Framework for analyzing low depth next-generation sequencing (NGS) data in heterogeneous populations using principal component analysis (PCA). Population structure is inferred to detect the number of significant principal components which is used to estimate individual allele frequencies using genotype dosages in a SVD model. The estimated individual allele frequencies are then used in an probabilistic framework to update the genotype dosages such that an updated set of individual allele frequencies can be estimated iteratively based on inferred population structure. A covariance matrix can be estimated using the updated prior information of the estimated individual allele frequencies.

The estimated individual allele frequencies and principal components can be used as prior knowledge in other probabilistic methods based on a same Bayesian principle. PCAngsd can perform the following analyses:

• Covariance matrix
• Genotype calling
• Admixture
• Inbreeding coefficients (both per-individual and per-site)
• HWE test
• Genome selection scan
• Kinship matrix

The older version, based on the Numba library (only working with Python 2.7) is still available in version 0.973 and can be found here [2].

Download and Installation

PCAngsd should work on all platforms meeting the requirements but server-side usage is highly recommended. Installation has only been tested on Linux systems.

Get PCAngsd and build

git clone https://github.com/Rosemeis/pcangsd.git
cd pcangsd/
python setup.py build_ext --inplace

Install dependencies:

The required set of Python packages are easily installed using the pip command and the requirements.txt file included in the pcangsd folder.

pip install --user -r requirements.txt

Quick start

# Download the input beagle file with genotype likelihoods
wget popgen.dk/software/download/NGSadmix/data/input.gz 


# Only estimate covariance matrix using 10 threads
python pcangsd.py -beagle input.gz -o test1 -threads 10

# Estimate covariance matrix and individual admixture proportions
python pcangsd.py -beagle input.gz -admix -o test2 -threads 10

# Estimate covariance matrix and inbreeding coefficients
python pcangsd.py -beagle input.gz -inbreed 2 -o test3 -threads 10

# Estimate covariance matrix and perform selection scan
python pcangsd.py -beagle input.gz -selection -o test4 -threads 10

Detailed Examples and Tutorial

Please refer to the tutorial's page [3]

Input

The only input PCAngsd needs and accepts are genotype likelihoods in Beagle format. New functionality for using PLINK files has been added (version 0.9). Genotypes are automatically converted into a genotype likelihood matrix where the user can incorporate an error model.

ANGSD can be easily be used to compute genotype likelihoods and output them in the required Beagle format.

./angsd -GL 2 -out data -nThreads 10 -doGlf 2 -doMajorMinor 1 -doMaf 2 -SNP_pval 1e-6 -bam bam.filelist

See ANGSD for more information on how to compute the genotype likelihoods and call SNPs.

Output

Since version 0.98, PCAngsd's output is only in binary Numpy format (.npy) except for the covariance matrix.

In order to read files in python:

import numpy as np
C = np.genfromtxt("output.cov") # Reads in estimated covariance matrix
S = np.load("output.selection.npy") # Reads results from selection scan

R can also read Numpy matrices using the "RcppCNPy" library:

library(RcppCNPy)
C <- as.matrix(read.table("output.cov")) # Reads in estimated covariance matrix
S <- npyLoad("output.selection.npy") # Reads results from selection scan

Using PCAngsd

All the different options in PCAngsd are listed here. PCAngsd will always compute the covariance matrix, where it uses principal components to estimate individual allele frequencies in an iterative procedure. The estimated individual allele frequencies will then be used in any of the other specified options of PCAngsd.´

# See all options in PCAngsd
python pcangsd.py -h

Estimation of individual allele frequencies

-beagle [Beagle filename]

Input file of genotype likelihoods in Beagle format (.beagle.gz).

-plink [Prefix for binary PLINK files]

Path to PLINK files using their prefix (.bed, .bim, .fam).

-plink_error [float]

Incorporate error model for PLINK genotypes.

-minMaf [float]

Minimum minor allele frequency threshold. (Default: 0.05)

-iter [int]

Maximum number of iterations for estimation of individual allele frequencies (Default: 100).

-tole [float]

Tolerance value for update in estimation of individual allele frequencies (Default: 1e-5).

-maf_iter [int]

Maximum number of EM iterations for computing the population allele frequencies (Default: 100).

-maf_tole [float]

Tolerance value in EM algorithm for population allele frequencies estimation (Default: 1e-4).

-e [int]

Manually select the number of eigenvalues to use in the modelling of individual allele frequencies (Default: Automatically tested using MAP test).

-o [prefix]

Set the prefix for all output files created by PCAngsd (Default: "pcangsd").

-indf_save

Choose to save estimated individual allele frequencies (Binary). Numpy format (.npy).

-dosage_save

Choose to save estimated genotype dosages (Binary). Numpy format (.npy).

-sites_save

Choose to save the marker IDs after performing filtering using population allele frequencies. Useful for especially selection scans and per-site inbreeding coefficients.

-post_save

Choose to save the posterior genotype probabilities. Beagle format (.beagle).

-threads [int]

Specify the number of thread(s) to use (Default: 1).

Call genotypes

Genotypes can be called from posterior genotype probabilities incorporating the individual allele frequencies as prior information.

-geno [float]

Call genotypes with defined threshold.

-genoInbreed [float]

Call genotypes with defined threshold also taking inbreeding into account. -inbreed [int] is required, since individual inbreeding coefficients must have been estimated prior to calling genotypes using that information.

Admixture

Individual admixture proportions and population-specific allele frequencies can be estimated based on assuming K ancestral populations using an accelerated mini-batch NMF method.

-admix

Toggles admixture estimations. Individual ancestry proportions are saved (Binary). Numpy format (.npy).

-admix_alpha [float-list]

Specify alpha (sparseness regularization parameter). Can be specified as a sequence to try several alpha's in a single run (Default: 0).

-admix_auto [float]

Enable automatic search for optimal alpha using likelihood measure, by giving soft upper search bound of alpha.

-admix_seed [int-list]

Specify seed for random initializations of factor matrices in admixture estimations. Can be specified as a sequence to try several different seeds in a single run.

-admix_K [int]

Not recommended. Manually specify the number of ancestral populations to use in admixture estimations (overrides number chosen from -e). Structure explained by individual allele frequencies may therefore not reflect the manually chosen K. It is recommended to adjust -e instead of -admix_K.

-admix_iter [int]

Maximum number of iterations for admixture estimations using NMF. (Default: 200)

-admix_tole [float]

Tolerance value for update in admixture estimations using NMF. (Default: 1e-5)

-admix_batch [int]

Specify the number of batches to use in NMF method. (Default: 10)

-admix_save

Choose to save the population-specific allele frequencies (Binary). Numpy format (.npy).

Inbreeding

Per-individual inbreeding coefficients incorporating population structure can be computed using three different methods. However, -inbreed 1 is recommended for low depth cases.

-inbreed 1

Simple estimator computed by an EM algorithm. Allows for F-values between -1 and 1. Based on ngsF.

-inbreed 2

A maximum likelihood estimator also computed by an EM algorithm. Only allows for F-values between 0 and 1. Based on [4].

-inbreed 3

Estimator using an estimated kinship matrix. Allows for F-values between -1 and 1. Based on PC-Relate.

-inbreed_iter [int]

Maximum number of iterations for the EM algorithm methods. (Default: 200)

-inbreed_tole [float]

Tolerance value for the EM algorithms for inbreeding coefficients estimation. (Default: 1e-4) Per-site inbreeding coefficients incorporating population structure alongside likehood ratio tests for HWE can be computed as follows:

-inbreedSites

Use likelihood ratio tests (.lrt.sites.gz) generated from -inbreedSites to filter out variable sites using a given threshold for HWE test p-value:

-hwe [LRT filename]

-hwe_tole [float]

Tolerance value for HWE test. (Default: 1e-6)

Selection

A genome selection scan can be computed based on posterior expectations of the genotypes (genotype dosages):

-selection

Using an extended model of FastPCA. Performs a genome selection scan along all significant PCs. Outputs the selection statistics and must be converted to p-values by user. Each column reflect the selection statistics along a tested PC and they are χ²-distributed with 1 degree of freedom.

Relatedness

Estimate kinship matrix based on method Based on PC-Relate:

-kinship

Automatically estimated if -inbreed 3 has been selected.

Remove related individuals based on kinhsip matrix of previous run:

-relate [Kinship filename]

-relate_tole [float]

Threshold for kinship coefficients for removing individuals (Default: 0.0625).

Citation

Our methods for inferring population structure have been published in GENETICS:

Inferring Population Structure and Admixture Proportions in Low Depth NGS Data

Our method for testing for HWE in structured populations has been published in Molecular Ecology Resources:

Testing for Hardy‐Weinberg Equilibrium in Structured Populations using Genotype or Low‐Depth NGS Data