veckit

Local scorer for the Virtual Embryo Challenge: score a submitted .h5ad against the T1/T2/T3 metric panels, entirely offline, against reference files you supply. This package never reads the official (held-out) validation/test data, and never imports anything that names where that data lives -- it only knows how to compute a metric panel from arrays you hand it.

pip install veckit

Register and download the released T1/T2/T3 stages from the official challenge site: virtualembryo.ai/challenge/data (see also aristoteleo/virtualembryo). veckit_tutorial.ipynb in this repo also ships a few tiny (150-cell) samples if you just want to try the tool first.

Full, meaningful panel (a real prediction, scored against the real preceding stage):

veckit --task T1 --input pred.h5ad --target T1/9.5.h5ad --reference T1/8.5.h5ad
veckit --task T2 --setting heart --input pred.h5ad --target b.h5ad --reference a.h5ad
veckit --task T3 --input pred.h5ad --target mab21l2_ko.h5ad --wt wt.h5ad

Quick copy_last / wt_identity check (--input doubles as --reference/--wt when omitted):

veckit --task T1 --input T1/8.5.h5ad --target T1/9.5.h5ad

from veckit import score
result = score(task="T1", input="pred.h5ad", target="T1/9.5.h5ad", reference="T1/8.5.h5ad")
print(result["metrics"])

--target (all tasks) is the pseudo target your --input prediction is scored against. --reference (T1/T2) / --wt (T3) is the reference expression de_score/de_direction/severity_slope are computed relative to -- these are PRIMARY metrics ("did you predict the right change", not just "does this look plausible"), so pass a real one whenever you're scoring a real model; omitting it defaults to --input itself, which is only correct when you're deliberately testing a no-change baseline (copy_last/wt_identity) -- those metrics then correctly read as exactly 0, matching the official baseline tables.

This is not a preview of your real competition score. Whatever you pass as --target is, by definition, data you already had — so this only tells you the scoring pipeline runs and your submission format is valid, not how well you'll do on the real held-out target.

Full task definitions and metric rationale: virtualembryo.ai and aristoteleo/virtualembryo.