Also thanks to 20% contributors: Ben Soudry, Mike Kruskal, Sowmiya Nagarajan, Suchismita Tripathy, Francisco Unda, Vasiliy Strelnikov

And community contributions from Sam Yadav and Seraj Ahmad at Roche improving the code for custom model training

Release led by Kishwar Shafin, contributions by Daniel Cook, Alexey Kolesnikov, Lucas Brambrink, and Pi-Chuan Chang as engineering manager.

Thanks to student researcher contributions from Farica Zhuang and Mobin Asri.

DeepSomatic release page:
github.com/google/deeps...

You have some additional control on memory use by the number of threads you run with.

For running on GPU, I am not sure if you've seen this - github.com/google/deepv...

Which requires a little more configuration, but can let you better manage CPU-GPU tradeoffs. Definitely expert use.

Hi Eric, sorry to not notice till now. From the DV FAQ, we see the Keras model takes 16GB of memory (github.com/google/deepv...).

It's possible that pangenome-aware models will take more memory, and we do observe more memory per thread used for that. Definitely not lower than 16GB.

Great question. We were talking recently about L40S benchmarks. We don't have that data immediately on hand, but are planning to generate runtime stats for it.

They're very close - to the point that small changes of coverage or the inclusion of PCR in preparation would tip between one and the other.

Release led by DeepVariant tech lead Kishwar Shafin. Team Engineering manager Pi-Chuan Chang. Small model work led by Lucas Brambrink. Pangenome-aware led by Mobin Asri and Juan Carlos Mier. Fast pipeline by Alexey Kolesnikov. Kinnex/MAS-Seq model by Daniel Cook and Shiyi Yin from Verily. 3/3

Added SPRQ to PacBio training, reducing Indel error on SPRQ by 26%. Added Platinum Pedigree training data for PacBio model, reducing errors by 34% on more extensive Platinum truth. New model and case study for Kinnex/Mas-Seq/Iso-Seq. Additional speed options for GPU pipelines 2/3

Release by Kishwar Shafin

Major contributions from Pi-Chuan Chang, Daniel Cook, Alexey Kolesnikov

Google 20%ers: Will Kwan, Pauline Sho, Lucas Brambrink, Mo Samman, Atilla Kiraly

UCSC for vg: @benedictpaten.bsky.social, Shloka Negi, Jimin Park, Mobin Asri

Pacbio: Billy Rowell, Nathaniel Echols

There are now custom models and case studies for CompleteGenomics instruments.

T7: github.com/google/deepv...

G400: github.com/google/deepv...

For now, these are stand alone models. We'll likely consider whether we can jointly include these in the broad WGS model later.

The changes to DeepTrio for de novo detection are substantial. We now in two steps - first for overall accuracy and then a weighted fine tuning for de novos. Our benchmarks show large improvements in de novo calling relative to the prior DeepTrio.

github.com/google/deepv...

Want to benefit from pangenomes and want a recipe?

github.com/google/deepv...

Shows a step by step process, with Docker images for how to map to a Pangenome reference w/ vg and calls w/ DeepVariant. Final calls are more accurate and in GRCh38 coordinates. Thanks to the UCSC team for co-development

Proud of Ayse Keskus and Asher Bryant in my group for making this happen! This work is a collaboration with Children's Mercy, UCSC and Google Health - who are also releasing the first version of DeepSomatic today: github.com/google/deeps...

TThanks to Jimin Park and Benedict Paten from UCSC, Mikhail Kolmogorov from NCI for analysis and testing. This group has also developed Severus for somatic SV calling, and we've worked closely with them

(github.com/KolmogorovLa...)

Thanks to Khi Pin Chua and Billy Rowell from PacBio