Reference rows

Returns the final training stage’s cells as the prediction. On Task 2 it carries that stage’s 3D coordinates across too, so the geometry is the previous embryo’s.

This is the floor: the score a model gets for changing nothing. It anchors the scale at 50, and it is not a weak opponent. Fitting a constant-velocity extrapolation off the evaluation stages lands below it on every setting. On the interpolation boards it is handicapped in a specific way worth knowing: the stage it copies is roughly 5.5× too small by volume, which the shape metrics penalise heavily.

Returns the target stage’s wild-type cells unchanged, keeping their coordinates.

Task 3’s floor. It is a strong one by construction: absolute expression barely moves after a knockout, so doing nothing already scores 0.956 absolute pseudobulk Pearson. That is exactly why the panel scores the wild-type-to-mutant change rather than the absolute state. Against this floor, absolute accuracy is not evidence of anything.

The target’s cells are split in two disjoint halves at a fixed seed. One half becomes the truth every submission is scored against; the other is submitted as a prediction.

It answers "how well can anything do here?": the ceiling is set by biological and sampling variation within the same embryo, not by an ideal model. Deliberately not the target resubmitted verbatim, which is a number no honest generative model could reach. Because it is an estimate from one half, a submission can beat it; the scale clips at 100.

For every cell type, take the difference of its mean expression between the last two training stages, Δ_c = mean(last|c) − mean(previous|c). Anchor on the real last-stage cells and add each cell’s own type’s Δ, clipped at zero. Types absent from the earlier stage stay where they are. On Task 2 the coordinates are again carried over unchanged.

The simplest thing that is not "do nothing": constant velocity, per cell type. It keeps within-type heterogeneity instead of collapsing to a mean, which is what makes it a fair reference rather than a straw man. It needs two visible stages to fit a difference, so it is absent from boards that have only one. Heart interpolation, for instance.

Measure the pseudobulk shift on the training knockout, Δ = pseudobulk(KO_train) − pseudobulk(WT_train), then add it to every cell of the target stage’s wild type, clipped at zero.

The paper’s named Task 3 baseline, and it encodes precisely the assumption the held-out perturbations exist to test: that knockouts share a common response. On the published references every model, this one included, gets the direction of the response wrong, so beating it is not a matter of a small architectural tweak.

Runs shift_transfer, then sets the knocked-out gene’s expression to zero in every predicted cell.

Isolates how much of the score comes from the single fact that the perturbed gene is gone. The gap between this and shift_transfer is small, which is the point: knowing which gene was removed is not the same as knowing what its removal does.

ref_X × 2.0, cells and coordinates otherwise untouched.

A metric insensitive to overall scale would score this as well as the honest reference. It is kept on the board so that if a future change reopens that hole, the row moves in public.

ref_X × 0.99.

The opposite test to ctrl_scale_ref: a metric that rewards any perturbation at all would score this above the floor for a change of one percent.

The reference stage’s pseudobulk mean, tiled to the same number of cells.

Scores respectably on any metric that only compares means, and near zero on the two that compare populations. It is the row that shows why half of Task 1’s weight sits on distributional questions, and it is the trap named explicitly in the submission notes.

Reference expression kept; coordinates redrawn uniformly at random between the true minimum and maximum along each axis.

A structureless cloud that merely fills the right volume. This is the control that caught a real hole: it beat the floor on the shape group at both interpolation boards, because copy_last is forced there onto a ~5.5× too small embryo while a right-extent blob is not penalised for size. The structure gate, shape credit multiplied by neighbourhood-structure skill, is what closed it.

Coordinates centred, multiplied by (4.0, 1.0, 0.25), then recentred. Expression untouched.

Preserves every local neighbourhood relationship while destroying the overall form. A shape metric that only looked at local structure would miss it entirely.

wt_X × 2.0.

The Task 3 analogue of ctrl_scale_ref: a large change in the wrong currency. The panel scores the change relative to wild type, so a uniform rescale should not read as a perturbation response.

wt_X × 0.75.

A smaller uniform change in the same spirit. Enough to move most absolute measures, in a direction that has nothing to do with the knockout.

A random Gaussian vector is rescaled to the same norm as the true wild-type-to-knockout shift, then added to the wild type and clipped at zero.

Separates magnitude from direction. Any metric that credits this is measuring only that something changed by about the right amount, which is why the panel scores the direction of the response explicitly, and why every published model getting that direction wrong is a finding rather than a rounding error.