Is there a way to identify mixed samples based on SNPs?

Example input (table of genotypes for multiple samples):

|      | sample1 | sample2 | sample3 |
| rs1  | AA      | BB      | AB      |
| rs2  | AA      | BB      | AB      |
| rs3  | AA      | BB      | AB      |
| rs4  | AA      | BB      | AB      |

Desired output: Sample relationships/distances. In this example, sample1 and sample2 are distinct, but sample3 is likely a mix of two (biological or technical).

Based on this simple example, the calculation seems doable, but maybe it's impossible in more complex realistic scenarios.

There is a nice tool peddy that does something conceptually similar, but it expects pedigree information as well as genotype frequencies for both the input samples and population.

  • $\begingroup$ I updated the question. I didn't initially realize it was not clear. $\endgroup$
    – burger
    Sep 15, 2018 at 1:48
  • $\begingroup$ Thanks, but this still remains a question that is mostly without context. What is the specific problem that you are trying to solve? As presented here, those samples could be individuals and the problem is a phasing problem, rather than mixed groups. What platform is being used to carry out the genotyping? How far away from each other are the variants? $\endgroup$
    – gringer
    Sep 15, 2018 at 5:19

1 Answer 1


Yes. The general term for identifying subsets from a mixed group is deconvolution.

This is easier to do when there is haplotype-level information available (e.g. sequenced reads that span multiple SNPs), but it is possible to get some indication of sample mixes from unlinked SNPs alone.

Deconvolution is helped a lot by having proportional data in the input (e.g. frequencies of genotypes, rather than a single genotype). Most SNPchip platforms have multiple probes per variant, and can provide some indication of variant proportion when digging down into the raw data. Deconvolution is also a lot easier with some knowledge about the component mixes. It's much easier to simulate mixes than to guess at possible mixes.

Here's a paper about the application of deconvolution in forensics.

Here's a paper about a program called CIBERSORT which carries out a deconvolution of immune populations based on gene counts:

Here's another paper which discusses the idea in more general terms.

You might find more information on one of these Biostars posts:

  • $\begingroup$ I am familiar with the expression-based deconvolution, but that's a different problem. And you can have the same mix of cell populations regardless of how many individuals are mixed in a single library. Do you have any suggestions for SNP-based deconvolution? $\endgroup$
    – burger
    Sep 14, 2018 at 19:29
  • $\begingroup$ Could you please clarify this in your question. It's not clear how much research you've done on this already, and whether or not you have haplotype-level information available. $\endgroup$
    – gringer
    Sep 14, 2018 at 23:12

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.