jrm5100

Yes, you could run different models and get different probabilities. For example, the likelihood that the sample is a mixture of the suspect, the victim, and some unknown person vs victim and two unknown people compared to saying the victim isn't in the sample. However, the specification of those models is part of the trial process.

And the output probabilities (at least when being used to determine guilt) are usually quite high, orders of magnitude higher than 90% or even 99.99%.

My point is that the science behind these calculations is well developed- validation studies get published all the time. Whether or not the specific software has errors (or isn't coded exactly as modeled) is an entirely different matter, but it still isn't all that likely. All of these cases rely on expert witnesses anyway- it's not the prosecutor pressing some buttons and printing a report.

There is far more concerning quackery that gets used in forensics- bite marks, hair matching, etc.

https://www.cancer.gov/about-cancer/causes-prevention/risk/d...

Fish cooked at high temperatures is also not good.

Monosopy.

This idea already here, but with technology based on older/simpler methods. One example is Intigenx: https://integenx.com/rapidhit-system/

The FBI has plans for implementing them: https://www.fbi.gov/news/testimony/fbis-plans-for-the-use-of.... The laws just need to catch up (or prevent it).

Minor nitpicking: the 23andMe kits use microarrays for genotyping, not DNA sequencing. It's an older and more limited technology but it is much cheaper than sequencing.

Public schools can use charters to game metrics themselves too [2].

[1] http://www.nber.org/papers/w22502

[2] https://www.propublica.org/article/alternative-education-usi...

The figure you linked is a good explanation. The split read method is helpful for finding the edges of the CNV, while the number of reads (relative to other regions that were tested) can give an idea of the number of copies. The problem is that these methods all have their own unique biases/noise that makes it non-trivial to figure out the absolute copy number change.

Ideally they would find a similar CNV that has some clinical association.

The DGV has a lot of reference CNVs. Here are some in BRCA1: http://dgv.tcag.ca/gb2/gbrowse/dgv2_hg19/?name=id:3087443;db...

"Copy number variant" refers to larger deletions and duplications that can occur in the genome. There isn't some specific cutoff for size, but some examples in these kinds of genes would be an entire exon or gene. There are countless studies that find correlations between specific variants or CNVs and risk of cancers.

Standard variant detection is pretty straightforward. CNVs are harder because they are longer (several hundred to several thousand base pairs) than the raw data (150 to 250 bp for Illumina)- you don't get single reads that span the entire variant. You have to normalize then look for differences in coverage, or look for split reads (where the read is aligned on the border of one of these CNVs).

This kind of funding baffles me because they don't seem to be proposing anything new at all (maybe slightly better CNV detection?) and there are already lots of labs/companies doing this kind of testing. Maybe they are working on being very efficient to offer a better price.