VCF-based expression predictions are slower than necessary because bcftools consensus is run against the full input VCF even when only a small region is specified.
I observed this clearly in tests/test_ad_risk.py and notebooks/vcf2exp.py, where the prediction took ~20 min using B200 GPU.
I have tested a simple fix (make small temporary VCF only for the specified region and apply it) and confirmed it reduces the same workload to ~2 min.
VCF-based expression predictions are slower than necessary because
bcftools consensusis run against the full input VCF even when only a small region is specified.I observed this clearly in
tests/test_ad_risk.pyandnotebooks/vcf2exp.py, where the prediction took ~20 min using B200 GPU.I have tested a simple fix (make small temporary VCF only for the specified region and apply it) and confirmed it reduces the same workload to ~2 min.