DeepVariant 1.0 (conference talk)

This is a presentation I gave in November 2020 at the (virtual) Biological Data Science meeting at Cold Spring Harbor Laboratory, based on my work on the Genomics team in Google Health, where I'm part of the team working on DeepVariant.
This covers how DeepVariant creates and classifies pileup images, some visual experiments we've done with it, and finally a look at the improvements in accuracy, runtime, and new models the team has made since DeepVariant was first open-sourced in 2017.

[@lethalfang]

Great presentation. Just the right amount of information for me!

[@jiajunfei354]

Great presentation! Thank you so much!

[@dBenedek]

Hi! Great video, thank you. I would be happy to hear from you more talks related to machine-learning in genomics. :)

[@YTARK8899]

Interesting sharing, Maria. It gives me an idea about how DeepVariant works. 👍

[@LiquidBrain]

Nice :) thanks for doing these

[@vladi1475S]

Thank you! :)

[@marbleous]

Hi Maria, awesome presentation and software! I'd like to ask how is the software giving variant results not found in ref and read (./.) and homozygous variants for both ref and reads ( 0/0), if the reference sequence is assembled from the reads? Thanks!

[@OMGenomics]

The reference genome is not a recently assembled genome from the same person as the sample we're now sequencing. The reference genome is an updated version of the one the Human Genome Project made, so it's the same one we use for all the analysis, and then when we sequence a bunch of new people, we compare them all back to that same reference genome. Does that help?

[@aliakay03]

Thank you so much for great explanation, I would like to ask that how deep variant handle, if the candidate variant in the last part of the read ?

[@aliakay03]

and what if we don't have 95 read in the pileup range?

[@OMGenomics]

Both of those are fine and normal. In the slides you can see most pileup images don’t fill out the full height with 95 reads. And there’s no reason why candidate variants toward the end of a read would be a problem.

[@yimingqu2403]

Hi Maria, I really love the work and how you present it. The flow is simply smooth and comfortable.
Do you think traditional tools like GATK is deprecated with all these new deep learning-based models proposed?

[@rukaiyavadnagarwala8041]

Hey! Maria! Can you throw sone light on OBAM, OBAMRC, OBP, OBF, OBQ, and OBQRC fields in a VCF file? These terms don't seem to be described in any documentation.

[@OMGenomics]

I don’t know anything about those fields either. Where did you see them?

[@rukaiyavadnagarwala8041]

@@OMGenomics I was asked to describe these fields as part of an internship task, could not find their meanings, so submitted it incomplete.

[@MGRVE]

Thank you. Very helpful. Where are the truth-datasets coming from?

[@OMGenomics]

Genome in a Bottle from NIST

[@MGRVE]

@@OMGenomics Thanks.