🤖 Written by Claude
SV annotation runs can take hours / never finish. Benchmarking (GRCh37/RefSeq/VEP 116, real prod data) shows the bottleneck is not the gnomAD-SV overlap (+3s) but the four conservation bigWig --custom summary_stats=max annotations (phastCons/phyloP), whose cost scales with SV span length:
- 1,486 SVs: baseline VEP 10.4s; +gnomAD 13.3s; +4 conservation bigwigs → timeout >290s
- single bigwig track: 559 tiny INS = 5.6s; 200 large SVs (0.5–4.5Mb) = timeout >200s
Proposed fix: drop the 4 conservation --custom bigwig args from the SV VEP command and compute the same _max columns in a separate pyBigWig stage. A validated PoC (aligned-window decomposition) is bit-exact vs VEP and does the 200 large SVs in 10.2s (VEP never finished).
Full write-up, benchmark tables, window convention, code sketch and integration steps: claude/sv_pybigwig_plan.md.
Implementation to be tackled later.
🤖 Written by Claude
SV annotation runs can take hours / never finish. Benchmarking (GRCh37/RefSeq/VEP 116, real prod data) shows the bottleneck is not the gnomAD-SV overlap (+3s) but the four conservation bigWig
--custom summary_stats=maxannotations (phastCons/phyloP), whose cost scales with SV span length:Proposed fix: drop the 4 conservation
--custombigwig args from the SV VEP command and compute the same_maxcolumns in a separate pyBigWig stage. A validated PoC (aligned-window decomposition) is bit-exact vs VEP and does the 200 large SVs in 10.2s (VEP never finished).Full write-up, benchmark tables, window convention, code sketch and integration steps:
claude/sv_pybigwig_plan.md.Implementation to be tackled later.