RNA-seq case study doc: mentions split_skip_reads=true is critical but example command doesn’t include it (and RNASEQ doesn’t auto-enable)

[brianjohnhaas]

Link: https://github.com/google/deepvariant/blob/r1.10/docs/deepvariant-rnaseq-case-study.md

Problem
In the RNA-seq case study, the “Flag summary” section states that split_skip_reads=true is Important and “critical for RNA-seq variant calling to work properly”, but the example run_deepvariant command shown above does not include any --make_examples_extra_args to set it.

This is confusing because it implies the option must be provided, yet the provided command omits it.

What I expected
Either:

The example command should include something like --make_examples_extra_args="split_skip_reads=true,channels=''" (exact quoting as needed), or
The documentation should explain that --model_type=RNASEQ automatically enables it (if that’s true).
What I found in code (r1.10)
scripts/run_deepvariant.py does not appear to automatically set split_skip_reads when --model_type=RNASEQ is used (no references to split_skip_reads in that script; only passes through --make_examples_extra_args if the user supplies it).
The split_skip_reads flag default is False in deepvariant/realigner/realigner.py.
References:

run_deepvariant.py (no split_skip_reads handling):
https://github.com/google/deepvariant/search?q=repo%3Agoogle%2Fdeepvariant+path%3Ascripts%2Frun_deepvariant.py+split_skip_reads&type=code
split_skip_reads flag default False:
https://github.com/google/deepvariant/blob/r1.10/deepvariant/realigner/realigner.py
Proposed fix
Update docs/deepvariant-rnaseq-case-study.md so the example command and the “Flag summary” are consistent:

Add the missing --make_examples_extra_args=... to the command, or
Adjust wording to clarify when/how the flag is enabled.

[brianjohnhaas]

my AI friend figured it out:

For DeepVariant 1.10.0, the RNASEQ model metadata does incorporate both settings.

I inspected the local container file directly:

{
"version": "1.10.0",
"shape": [100, 221, 6],
"channels": [1, 2, 3, 4, 5, 6],
"flags_for_calling": {
"split_skip_reads": true,
"pileup_image_width": 221,
"max_reads_per_partition": 0,
"partition_size": 10000,
"channel_list": "BASE_CHANNELS",
"min_mapping_quality": 40
}
}

That means:

• split_skip_reads=true is encoded in model.example_info.json.
• The channel configuration is also encoded there.
  For RNASEQ in 1.10.0, it is not literally channels=''; instead the metadata sets channel_list: "BASE_CHANNELS" and the model declares channels: [1,2,3,4,5,6].

I also verified from r1.10 run_deepvariant.py that:

• --model_type=RNASEQ maps to /opt/models/rnaseq
• it auto-discovers /opt/models/rnaseq/model.example_info.json
• it passes that file to make_examples via --checkpoint_json
• the code comment explicitly says that starting in v1.10.0, calling-time make_examples args come from model.example_info.json

[danielecook]

@brianjohnhaas thank you for spotting this. I will remove those flags to make the documentation clearer. DeepVariant will set flags based on a json file stored alongside models as you pointed out.