Using SVision to detect variations in tandem repeats

[CSU-KangHu]

Hello @jiadong324,

Thank you for developing such an excellent tool. I have used SVision to detect tandem repeat variations in the HG002 sample with the GRCh38 reference. I employed the default command provided in the demo section:

SVision -o ${pathTo}/run_svision -b ${pathTo}/HG002.m84011_220902_175841_s1.GRCh38.chr1.bam -m ./svision_model/svision-cnn-model.ckpt -g ${pathTo}/GCA_000001405.15_GRCh38_no_alt_analysis_set.fasta -n HG002 -s 5 --graph --qname

Then, I used Truvari to evaluate the results with the following commands:

truvari bench -b ${pathTo}/HG002_GRCh38_TandemRepeats_v1.0.chr1.vcf.gz -c ${pathTo}/HG002.svision.s5.graph.sorted.vcf.gz --includebed ${pathTo}/HG002_GRCh38_TandemRepeats_v1.0.chr1.bed.gz --sizemin 5 --pick ac -o bench_result/
truvari refine --use-original-vcfs --reference ${pathTo}/hg38_chr1.fa bench_result/

Where the files HG002_GRCh38_TandemRepeats_v1.0.chr1.vcf.gz and HG002_GRCh38_TandemRepeats_v1.0.chr1.bed.gz are gold standard tandem repeat variation files and regions downloaded from https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/release/AshkenazimTrio/HG002_NA24385_son/TandemRepeats_v1.0/GRCh38/.

The results are displayed in the figure below:

The recall is low. I have also checked some tandem repeat variation regions that were not detected by SVision. The IGV figure is shown below:

Many such regions were not detected by SVision. I wonder if SVision is not designed to detect tandem repeat variation regions?

I also performed the same test using TRGT, and the experimental results can be seen at ACEnglish/adotto#5.

[songbowang125]

Hi, I looked through your ground truth set (HG002_GRCh38_TandemRepeats_v1.0) and found that most of the variants were small variants, like indels smaller than 50bp. Also, the variant in your your example IGV was also a small variant. Since SVision detect (complex) structural variants rather than small variants.