50bp deletion not detected

[HenrivdGeest]

Im strugling with clair3 to get a ~50bp deletion detected. Any other variant I can see by eye is detected, but I have a deletion in ~ 100% of my reads, but for clair3 it seems hard to find.

I use the 0.1-r8 release, with the following parameters:
/opt/bin/run_clair3.sh --bam_fn ${NEWBASE}.bam --ref_fn ${reference} --threads=4 --platform=ont --model_path="/opt/models/ont"
--output ${NEWBASE}_clair3 --include_all_ctgs --snp_min_af=0.01 --indel_min_af=0.001
My data is ONT, pcr amplicon reads, >1000x coverage, downsampled to ~300x.
I've tested a few subsamplings, and in 1 case at 50x coverage it did find the deletion, but in another random sampling to again 50x, it did not find it.
Is there something I can do about this, or tweak some parameter so that it does find it?

[aquaskyline]

Clair3 reports SNP and indels, and it follows the definition that indels are <50bp (≥50bp are structural variants). A hardcoded threshold in Clair3 limits the longest indels to be reported (https://github.com/HKU-BAL/Clair3/blob/main/clair3/CallVariants.py#L27). Elevating the threshold enables Clair3 to report insertions and deletions ≥50bp. But we have not tested Clair3's performance on indels ≥50bp, it very much depends on the read length and the performance of the sequence aligner.

[HenrivdGeest]

I see, that makes sense. I will try a SV caller for this example. Thanks!

[aquaskyline]

No. Actually, you reminded me to question the rationality of using 50bp as the indel size cutoff in the ONT era. The 50bp cutoff was to a large extent a practical cutoff imposed by the short length of NGS reads. Before NGS, the cutoff was 1kbp. I will be extending the limit to a larger value say maybe somewhere between 200bp to 500bp depending on the maximum reliable length of an opened gap in a typical length ONT read.

Indel length cutoff elevation scheduled for v0.1-r9, stay tuned.

[aquaskyline]

Please try out the latest version v0.1-r9 with option --enable_long_indel to see if the 50bp deletion in interest could be identified.