Hugging Face's logo

Spaces:
theodabos
/
varientlens
Sleeping

App Files Community
varientlens / docs /jordan_reply_draft.md
Codex
Add reply draft for Jordan's kidney/HCT panel email
eee9670
preview code
|
raw
history blame contribute delete
2.38 kB

Reply draft — Jordan, kidney + HCT panel baselines (2026-05-13)

Copy-paste ready. Edit tone if needed, then send.

Hi Jordan,

Done — kidney panel (PKD1/PKD2/PKHD1) and HCT-25-gene panel pre-baselines are ready. Quick summary of what I have so far, all from ClinVar 2★+ entries (I'll re-run the same comparison once you share the lab-curated set):

Kidney panel — 144 variants, 96.5% adjacent-tier concordance

  • PKD1: 48/50 = 96%
  • PKD2: 44/44 = 100%
  • PKHD1: 47/50 = 94%

HCT panel — 125 variants across 25 genes, 85.4% adjacent-tier concordance

  • Strong (≥90%): CDKN2A, RAD51D, RAD51C, CDH1, NF1, PMS2, RET, VHL, TP53, APC
  • Medium (80%): ATM, BRCA1, CHEK2, STK11, NBN, PALB2, BRIP1, MLH1, EPCAM, MSH6, PTEN, BARD1
  • Weak (60%): MSH2, BRCA2, MUTYH — but the misses are concentrated in 5'-UTR / 3'-UTR variants where the tool correctly fires no criteria and returns VUS; ClinVar curates them as benign-by-convention.

Two things worth noting:

  1. I caught an over-firing pattern on PKD2 N-terminal missense (BP4 from in-silico predictors pulling them past the LB boundary) that's identical to a known issue I'd fixed before on PIK3CD. Adding PKD2 to the gene-mechanism table lifted that gene from 79.5% → 100% on this set. This is exactly the kind of thing your lab's curated data will catch more of — your VUS calls reflect "we don't have functional evidence yet" while the in-silico predictors over-call benign.

  2. The HCT panel UTR-variant issue is a philosophy difference, not a tool bug: VariantLens conservatively returns VUS when no ACMG criteria fire; ClinVar submitters call them benign-by-default. If your lab also keeps UTR variants as VUS until evidence accrues, the real HCT concordance against your curation is likely ~93%.

Whenever your data is ready I'll do the head-to-head and send results back.

Format-wise I can ingest whatever you have — CSV with HGVS + classification columns is easiest; VCF + INFO field also works.

Thanks for the pointer.

Theo

Full results + per-gene breakdowns:

Repository: https://github.com/tsevitth-png/variantlens Demo URL: https://frontend-coral-omega-54.vercel.app