This is a test version of Biostars. For the public version, visit https://www.biostars.org.
One-hot encoding for PLINK or VCF

I want to write an autoencoder for SNP data. Is there an established way to one-hot-encode binary PLINK or VCF input? I believe that can be done by manipulating PLINK's bed file but am afraid to do something wrong.

By one-hot encoding I mean MISSING = [1000] HOM_REF = [0100] HET = [0010] HOM_ALT = [0001]

Thanks!

one-hot-encoding plink vcf
• 2,545 views • link
updated by carlk • carlk • written by dmitry.s.kolobkov • dmitry.s.kolobkov

why 4 bytes? PLINK's internal uses 2 byte per genotype, 00, 01, 11, and 10. Can't remember if 01 or 10 is missing.

• 0 views • link
written by Sam Sam

yes but I need to use it to train a neural network so I need four separate variables for four categories of a single SNP

• 0 views • link
written by dmitry.s.kolobkov • dmitry.s.kolobkov

I see. Then I don't know if there's anything like that. Out of interest, can't you also represent the same data with 000, 001, 010 and 100? Though I guess being a multiple of 2 might be more efficient.

• 0 views • link
written by Sam Sam

1 answer

Here is what we do for FaSTLMM, which does linear mixed models, a generalization of linear regression. We treat the values as 0,1,2 (the count of the A1 alleles), then we standardize (for each variant, set mean to 0 and s.d. to 1). Finally, we set any missing values to 0.

Put another way, we treat the values as real-valued input, not as two binary inputs. I think these steps are common in the field. It was my impression that most neural nets are happy with real-valued input.

If you're working in Python, we have created bed-reader for reading PLINK bed files. Under the covers, it uses a fast multi-threaded Rust engine. It supports all Python indexing methods and you can slice data by individual (samples) and/or by variants.

  • Carl
• 0 views • link
written by carlk • carlk

Log in to answer this question.

More posts like this