10 ms·

DeepVariant didn’t actually explicitly encode the variants (nor the raw data) as images. Press reports —including from Google Research itself) suggested that th

by klmr 8y ago

DeepVariant didn’t actually explicitly encode the variants (nor the raw data) as images. Press reports —including from Google Research itself) suggested that this was the case but nothing in the original publication said so, and the researchers themselves have disputed it. It just so happens that the tensor representation lends itself well to visualisation as (multi-channel) images. But that wasn’t the intent, it’s just a nice side-effect. In reality the data is laid out in tensors that follow naturally from how they were generated (i.e. via alignment of many stochastically generated DNA fragments to a reference sequence).

a_bonobo 8y ago
Interesting!! I did not know that, so I was wrong - but why does the GitHub repo then talk of pileup images? https://github.com/google/deepvariant https://github.com/google/deepvariant Edit: I think I get it now, this notebook shows it: https://github.com/google/deepvariant/blob/r0.7/docs/visualizing_examples.ipynb https://github.com/google/deepvariant/blob/r0.7/docs/visuali... The channels are easily representable as images, but the network doesn't use the images directly.
- klmr 8y ago
  Yeah, the way they talk about this is admittedly confusing. The best way for me to think about it (from [1]) is to mentally add quotation marks around the word “image” whenever the methods mention it. Mathematically the data is stored in higher-dimensional tensors with one dimension representing genomic coordinates, another dimension representing the sequence depth (i.e. one row per sequence read), and additional dimensions representing features of the sequence (such as nucleoside identity, read direction, error probability, match/mismatch with the reference, etc). I use almost the same kinds of tensors for work that has no tangible relation to images or visualisation. It’s simply the most straightforward way of representing this data as a tensor. But the similarity to images is so tantalising that even the paper’s methods [2] fall prey to this, even though the dimensionality and minor details are wrong. Furthermore, the term “pileup image” refers to a common way of visualising genome–read alignments [3]. The DeepVariant tensor is not a pileup image but it is very close. And the tensor can be converted in to an image [4] but as mentioned this requires some transformations (splitting the channels, and rescaling the values). [1] https://bioinformatics.stackexchange.com/q/4098/29 https://bioinformatics.stackexchange.com/q/4098/29 [2] I initially claimed the paper didn’t mention this. That’s wrong; apologies. [3] https://www.google.com/search?tbm=isch&q=genome+pileup https://www.google.com/search?tbm=isch&q=genome+pileup [4] https://github.com/google/deepvariant/blob/r0.7/docs/visualizing_examples.ipynb https://github.com/google/deepvariant/blob/r0.7/docs/visuali...
  - a_bonobo 8y ago
    Thank you for the explanation, it makes perfect sense - as an RGB picture is represented in a three dimensional list of numbers for red, green, blue you're not constrained by that, you might as well throw anything in there.

by klmr 8y ago

a_bonobo 8y ago

Interesting!! I did not know that, so I was wrong - but why does the GitHub repo then talk of pileup images? https://github.com/google/deepvariant https://github.com/google/deepvariant Edit: I think I get it now, this notebook shows it: https://github.com/google/deepvariant/blob/r0.7/docs/visualizing_examples.ipynb https://github.com/google/deepvariant/blob/r0.7/docs/visuali... The channels are easily representable as images, but the network doesn't use the images directly.

klmr 8y ago
Yeah, the way they talk about this is admittedly confusing. The best way for me to think about it (from [1]) is to mentally add quotation marks around the word “image” whenever the methods mention it. Mathematically the data is stored in higher-dimensional tensors with one dimension representing genomic coordinates, another dimension representing the sequence depth (i.e. one row per sequence read), and additional dimensions representing features of the sequence (such as nucleoside identity, read direction, error probability, match/mismatch with the reference, etc). I use almost the same kinds of tensors for work that has no tangible relation to images or visualisation. It’s simply the most straightforward way of representing this data as a tensor. But the similarity to images is so tantalising that even the paper’s methods [2] fall prey to this, even though the dimensionality and minor details are wrong. Furthermore, the term “pileup image” refers to a common way of visualising genome–read alignments [3]. The DeepVariant tensor is not a pileup image but it is very close. And the tensor can be converted in to an image [4] but as mentioned this requires some transformations (splitting the channels, and rescaling the values). [1] https://bioinformatics.stackexchange.com/q/4098/29 https://bioinformatics.stackexchange.com/q/4098/29 [2] I initially claimed the paper didn’t mention this. That’s wrong; apologies. [3] https://www.google.com/search?tbm=isch&q=genome+pileup https://www.google.com/search?tbm=isch&q=genome+pileup [4] https://github.com/google/deepvariant/blob/r0.7/docs/visualizing_examples.ipynb https://github.com/google/deepvariant/blob/r0.7/docs/visuali...
- a_bonobo 8y ago
  Thank you for the explanation, it makes perfect sense - as an RGB picture is represented in a three dimensional list of numbers for red, green, blue you're not constrained by that, you might as well throw anything in there.