Three tasks, one shared resource

Ground truth is withheld while it can still affect the board. Until the final phase a validation submission returns a leaderboard score, not the answers. At the start of the final phase the validation answers are released for every task — ranking has moved to the hidden test split by then, so they can no longer be used to probe it — and the test leaderboard opens. Test ground truth is never distributed.

T1Temporal

Temporal gene-expression distribution prediction

Given stages before a target time, predict the gene-expression distribution — a set of cells — at a future stage. Tests temporal extrapolation.

Modality: single-cell RNA
Predicts: expression
Splits: Models train on E8.5 and E9.5, are scored against E10.5 through the leaderboard, and are ranked on a hidden E12.5.
Scored on: DE gene recovery (25%), Change direction (25%), Cell-state distribution (30%), Gene-gene co-variation (20%)
Why it’s hard: Nothing is observed between E10.5 and E12.5, so there is no bracketing stage to interpolate from. And the target is mostly not the training population moved forward: E8.5 and E9.5 share only 5 cell types, so about two thirds of the cells at the later stage are types that were not there before. They did not drift into place, they differentiated — which makes this a question about which cells appear, not about how far existing ones travel.

Full task page →How it is scored →

T2Spatial-temporal

Spatial-temporal multiscale prediction

Predict future gene expression and 3D spatial location jointly — molecular, cellular and tissue scale at once. Scored separately on the heart and on the whole embryo.

Modality: 3D MERFISH (500-gene panel)
Predicts: expression + 3D coordinates
Splits: Interpolation is trained and validated on both heart and embryo, but only tested on embryo; extrapolation lives entirely in the heart setting. Heart, in order: E8.25 train, E8.5 interpolation validation, E8.75 train (also Task 3’s wild-type reference), E9.5 train, E10.5 extrapolation validation, E12.5 extrapolation test — and in the final phase every heart stage is training input. Embryo, in order: E6.75 train, E7.25 train, E7.5 validation, E7.75 test, E8.0 train — the embryo setting has no extrapolation target, E8.0 being its latest stage.
Scored on: Expression change (25%), Cell-state distribution (25%), Tissue shape and growth scale (25%), Local spatial organisation (25%)
Why it’s hard: Three scales are scored at once, which is what multiscale means here: molecular — does each gene move the way it moves in the real embryo; cellular — is the mix of cell states right, in the right proportions; and tissue — is the thing the right size and shape, with cells beside the neighbours they should have. Each carries a quarter of the score, so getting the marginals right is not enough on its own. The held-out stages also sit both inside the observed range and beyond it, so a model has to interpolate and extrapolate.

Full task page →How it is scored →

T3Perturbation

Mutant perturbation prediction

Predict a held-out knockout — expression and 3D coordinates — from wild-type development plus one observed perturbation.

Modality: 3D MERFISH (500-gene panel)
Predicts: expression + 3D coordinates
Splits: The Mab21l2 knockout at E9.5 is released to train on, Gata4 at E8.75 is the validation target, and β-catenin at E8.75 is the hidden test.
Scored on: Response gene recovery (30%), Response direction (25%), Response magnitude (25%), Cell-state distribution (20%)
Why it’s hard: A mutant embryo still looks mostly like a wild-type one: hand back the matched wild type unchanged and absolute expression already correlates at 0.956. So the score is computed on the difference between wild type and mutant, and a submission earns nothing for reproducing the wild type it was given — only for moving the genes the knockout actually moves.

Full task page →How it is scored →

Two parallel tracks

Both tracks address the same three tasks and are scored on the same metrics and hidden test sets. Prizes are awarded separately, so the leaderboards directly contrast human-designed and agent-designed approaches.

Track 1 Human Team: Conventional ML-competition workflow.
Methods designed and supervised by human participants. Algorithm/model design → submission → evaluation. Standard NeurIPS competition track.
Track 2 Agent Team: Coding agents / LLM-driven recursive systems.
Methods produced by coding agents or LLM-based evolutionary systems. A human may write the initial prompt; from there the run must be the agent’s own — no human inspecting intermediate results and feeding judgement back in. What counts is either carrying a published method through optimisation end to end, or inventing and implementing a new algorithm from scratch. Prizes require evidence: the trajectory, the prompts, and the harness code. What cannot be verified cannot win.