Common questions

Anyone — academic, industry, independent — who can submit a prediction file conforming to the required format. Each person joins one team; each team works on one track.

Yes. The three tasks share a data resource and a metric panel but are scored separately, so a team can enter one, two, or all three.

Yes — coding agents, recursive LLM systems, evolutionary search. To be eligible for prizes, the agent system and full evolutionary trace must be shared before prize evaluation.

They are scored on the same tasks, the same metrics and the same hidden test sets, and their leaderboards sit side by side — but prizes are awarded per track, so a human-designed method never competes with an agent-designed one for the same prize.

A predicted set of cells for the target condition: expression for Task 1, expression plus 3D coordinates for Tasks 2 and 3. You never submit cell-type labels — the organisers assign types with a frozen classifier applied identically to every submission, so hidden labels are never exposed.

No. Nothing in the metric panel requires matching cell counts, and where a comparison would otherwise be confounded by count — the shape group — both point clouds are subsampled to a shared size first. Cell number is scored separately and explicitly, as count_log_ratio.

No. Every spatial metric is agnostic to translation, rotation and — outside one acknowledged blind spot — reflection. That blind spot is laterality: a mirrored embryo currently scores identically to a correct one.

Baseline implementations for each task, data-loading utilities, the exact evaluation code, example submission files, a standalone scorer for any .h5ad, documentation, and executed notebook tutorials.

Training data is released for method development. Validation and test ground truth are both withheld: a validation submission returns a leaderboard score, and the answers are released only when the final test set is.

Yes for Task 1 — only the evaluation stages are fixed, and participants may additionally train on external public single-cell data. Any external source must be disclosed with the submission.

Hidden splits are defined by developmental stage, perturbation condition, or combinations thereof — distinct from public splits. Models must learn generalisable developmental dynamics, not memorise.

Two anchors published with every score. The floor is copy_last — predict the preceding stage verbatim (wt_identity for Task 3). The ceiling is half the target scored against its other, disjoint half — deliberately not the target resubmitted verbatim, which is a number no honest generative model can reach.

A metric your output cannot produce, or that comes back NaN, counts as zero skill within its group. It is never treated as an excused absence.

Every attack found during the adversarial review is kept as a control row that is re-scored on every run. If a future change reopens one of those holes, the control row scores above its floor and the scoreboard flags it — the regression shows up in the table instead of needing another audit to find.

All ground-truth labels are hidden. Submission counts are limited in the final phase to reduce leaderboard probing. Duplicate registrations, unauthorised data use, code sharing across teams, or falsified results are grounds for disqualification.