AutoZyme Lab
Lab tasks turn slow bioinformatics methods into reproducible optimization challenges. Download the frozen task card, run the public gate locally, and upload an evidence bundle. Public results move a patch into review; verified speedups are produced later by AutoZyme maintainers on hidden OOD data and credited to accepted contributors.
How a submission becomes a new round
Uploaded bundles enter a review queue first. AutoZyme checks the public evidence, reruns the candidate on a maintainer machine, and then evaluates a private held-out suite. Only submissions that improve verified speed while passing every frozen gate are appended to the trajectory as a new credited round and considered for packaging or upstream contribution.
seurat_findallmarkers_v1Seurat FindAllMarkers v1
Optimize Seurat's differential marker discovery while preserving marker sets, log-fold-change ranking, adjusted p-values, and output schema.
Downloads
Challenge freeze
seurat_findallmarkers_v1v186c567dedab1...6a796ee3Public smoke test
PBMC small smoke test · 80 cells · 230 genes
Rscript run_public.R --data public_smoke_pbmc_small.rds --impl solution_template.R --out run_public
Bundle contract
Include run_public/results.json, source changes, environment notes, and any risk notes needed for review.
Public dev status
PBMC68k public dev dataset · external mirror pending
Hidden OOD
Held-out single-cell datasets with larger scale, different tissue context, different sparsity, and different cluster structure. Exact datasets and reference outputs are not disclosed.
Optimization trajectory
FindAllMarkers public development history. Hover each round for the mechanism and submitter.
Submit a bundle
Submission contract
The upload API performs a public sanity check and stores your evidence bundle. Verified leaderboard numbers come only from maintainer reruns against the frozen task and hidden OOD suite.
- Expected bundle path:
run_public/results.json. - Include source changes, dependency changes, environment notes, and any approximation risks.
- Do not upload large intermediate outputs; include hashes or reproducible scripts instead.
- Accepted submissions are credited in the trajectory and may be packaged or proposed upstream.