DeepVariant on BGI / DNBSEQ PE150 data — model suitability and QC guidance

[Passion-Zoey]

Hi DeepVariant team and community,
Thanks for maintaining DeepVariant — it’s a very powerful and widely adopted variant caller.
We are working with "human whole‑genome sequencing (WGS) data generated on the BGI / DNBSEQ platform (PE150)". Although DNBSEQ produces short reads similar to Illumina platforms, the underlying sequencing chemistry and error characteristics are distinct. DeepVariant’s standard WGS model is primarily trained and benchmarked on "Illumina short‑read data" in existing publications and documentation. At the same time, we’ve also seen references to "Complete Genomics / DNBSEQ‑specific DeepVariant case studies (e.g., T7/G400)", suggesting the possibility of platform‑specific models.

Our Questions

1. For DNBSEQ PE150 human data, is the standard DeepVariant WGS model recommended, or should users prefer a Complete Genomics / DNBSEQ‑specific model if available?
2. Are the "T7/G400 case‑study models" mentioned in documentation or presentations publicly available for general use, or were they only for demonstration and internal benchmarking?
3. For older or externally service‑generated DNBSEQ datasets, what QC metrics would you recommend to assess whether the standard WGS model is acceptable? For example:
   • Ti:Tv ratio expectations
   • GQ calibration checks
   • Indel / SNP balance
   • Depth / mapping consistency
   • Platform‑specific patterns to watch for
     We want to ensure that our germline variant callset is robust and that the model choice is appropriate for the sequencing technology.
     Thanks very much for any guidance!

Best regards,
Yi Zhao

[AndrewCarroll]

Hi @Passion-Zoey

We generally observe that Illumina WGS models perform well with Complete data. It is viable to use the WGS model with complete, often exceeding the results when we run Illumina data with the Illumina model. We do observe a ~5-10% error reduction when using the Complete model with Complete data. We observe that T1 and T7 models are the most interchangeable with each other.

I am sharing the documentation we have for available models. Note that some have been trained for pangenome-mapped samples, while others trained for BWA. The most important thing to do is to match the model to the mapping choice.

T7+ model with pangenome-mapping: https://gist.github.com/pichuan/894bdae79bcf4bd93d1396d166c2425c

At this time, use this model for T1 or T7. We are writing up documentation for the pangenome-T1, but the T1/T7 models are fairly comparable.

The case studies you see here are for linear mapping (BWA). Use T7 for T1 or T7 data.

https://github.com/google/deepvariant/blob/r1.10/docs/deepvariant-complete-t7-case-study.md
https://github.com/google/deepvariant/blob/r1.10/docs/deepvariant-complete-g400-case-study.md

We intend to organize and make all model choices documented between now and the next release.

In terms of assessments, if you have sequenced Genome in a Bottle samples, the best way to assess a model is to run the model on that sample and look at the accuracy reports from hap.py.

Failing that, the best way to understand the difference between models is to run two models and look at the discordant variant calls between the two models, and use metrics such as the number of calls and Ti:Tv as an indirect measure.

Thank you,
Andrew

[Passion-Zoey]

Hi @AndrewCarroll
Thank you very much for your detailed explanation. We have confirmed with the sequencing provider that our samples were generated on the DNBSEQ-T7+ platform.
Our current pipeline uses bwa-mem2 for linear mapping to GRCh38, followed by DeepVariant v1.10.0.
Could you please advise whether it would be appropriate to use the Complete T7 linear/BWA customized model for T7+ data under this linear mapping workflow?
Thank you again for your help.
Best,
Yi Zhao

[AndrewCarroll]

Hi @Passion-Zoey

Yes, the Model described in this case study - https://github.com/google/deepvariant/blob/r1.10/docs/deepvariant-complete-t7-case-study.md

Is the best model to use for BWA-mem + T7 Compelte data.

Hope that definitively answers your question.