Hacker News Reader

Comment by nkrumm | Hacker News Reader

It's true that these companies are not offering "sequencing" per se but rather are genotyping commonly polymorphic SNPs using an array or other genotyping method.

These SNP arrays, however, are exceedingly accurate and reproducible in their calls. A common medium-density SNP array used for human genetics is the Illumina CytoSNP-850K [1], which routinely gets >99.5% accuracy on calls. This is at least as good, or better than, whole genome sequencing.

The challenge in ancestry assignment is not determining what is in the sample, but rather in the interpretation and assignment of ancestral haplotypes.

[1] https://www.illumina.com/content/dam/illumina-marketing/docu...

View on HN

Since the test results for the twins in the article were 99.6% identical, that seems to match what you say is the usual accuracy of SNP arrays. However, I don't think that can be called exceedingly accurate and reproducible if you consider the intended application.

>99.5% accuracy is great for a scientific measurement device that's only been developed very recently; it's absolutely rubbish if you want to sell the results to consumers.

If the practical difference is 38% vs 39% "Eastern European", who cares?

">99.5% accuracy is great for a scientific measurement device that's only been developed very recently; it's absolutely rubbish if you want to sell the results to consumers."

Or potentially put people in jail based on these results...

If you feel that way about 99.5% accuracy, I'd advise that there are a number of consumer products you absolutely shouldn't buy, including: kitchen scales (small ones), measuring cups, store brand spirit levels, etc.

You can get considerably better accuracy, you just have to be willing to fork over $1500 rather than $60. How much are you willing to pay to know if your 31% German vs 30%?

The difference is that it's not 99.5% accuracy for the entire experiment but for each of the millions of spots, which means there will be hundreds of thousands of spots that are not accurate in each chip. Not an apples to apples comparison.

I am inferring that the problem of matching a particular individual's results with haplotypes is a problem in optimization (finding a best fit) with many variables that is actually solved by maximization (finding a good fit) without doing a comprehensive search of all possible solutions. The space of potential results may be very large -- very many haplotypes from which to select and a percentage to be assigned to each, from 0 to 100, and a perfect result cannot be expected, as ancestors within a given haplotype are nowhere near as similar as identical twins. It is very common, even typical, for algorithms that attempt to find good solutions to such problems to include some randomization (making some guesses based on pseudo-random numbers) followed by more directed attempts to successively improve the guesses previously made. Because some steps of the search are a random walk, different results from the same inputs are common for several iterations of the same algorithm.

Reply on news.ycombinator.com