PCA results from PLINK and Hail vastly different

Question

[Ahmed]

I am getting completely different results when I conduct PCA on PLINK and on HAIL - does anyone know why? When I say the results are different I mean:

1. Comparing the pearson correlation between the top 10 PC's there is 0 correlation
2. When I create a PCA scatter plot I get completely different looking clusters suggesting different population stratification

Points to note:

1. Its the same set of samples and SNPs (I am using the same .bed/.bim/fam files)
2. I did QC on the dataset prior (including LD pruning, MAF > 0.05, genotype > 0.95). From the hail info none of the SNPs are being removed (it says the number of SNPS left after filtering is the same as I had in my .bim file)
3. When I use another software (bigsnpr) I get clusters close to what I get in Hail.

My commands are as follows:

HAIL v0.2

hl.import_plink(bed =file.bed, bim =file.bim,  fam =file.fam, reference_genome='GRCh38' ).write("file.mt', overwrite = True)
samples = hl.read_matrix_table('file.mt')
pca_evals_s, pca_scores_s, pca_loadings_s = hl.hwe_normalized_pca(samples.GT, k=10, compute_loadings=True)

PLINK2.0

plink2.0 --bfile file --pca 10 --out plink_pca --threads 14

EDIT

The issue only happens with plink2.0 and not with plink1.9

Thank you!

[LChart]

If you run plink2 --version what is the result?

Answer 1

[chrchang523]

As noted in the plink2-users Google group (https://groups.google.com/g/plink2-users/c/DeTVfXAjzTY ), this was due to use of a 6-year-old plink2 build.