The pancreas result reproduces directionally, but TopoMetry's 2-D layouts ignore their seed — so no single-run layout number should be quoted.

TopoMetry cell-cycle replication

What was tested

topometry builds geometry-aware spectral scaffolds for single-cell data. The claim under test is its developing-mouse-pancreas result: that TopoMetry’s representations arrange cycling ductal cells in cell-cycle-phase order more faithfully than conventional linear embeddings.

This is reproduction, not reimplementation. TopoMetry is MIT and the surrounding stack is BSD, so the published code was read, installed, and run as released. The dataset is 3,696 cells and the whole analysis runs in minutes.

What came out

The headline finding is about reproducibility, not geometry. TopoMetry’s spectral scaffolds are deterministic — three refits under identical parameters agree to a pairwise-distance correlation of 1.0000. Its 2-D layouts are not, despite random_state=0: the loop statistic on TopoMAP ranges 0.532 to 0.918 across identical runs.

Since the paper’s pancreas claim is a claim about a layout, no single-run layout number should be quoted without that spread — the paper’s included. A reported value could sit anywhere in that band, and a comparison between two methods can be smaller than the noise in one of them.

Within that limit the published observation reproduces directionally. Cycling ductal cells are arranged in phase order far more strongly in TopoMetry layouts (|ρ| = 0.897) than in linear PCA (0.620), and the single-time spectral scaffold beats PCA on all three of the paper’s own preservation metrics.

Two sub-claims did not reproduce here. The multiscale scaffold ties PCA rather than beating it, and TopoMetry’s 2-D layouts score below pca_umap on the preservation metrics. And TopoMetry is not shown to beat the conventional PCA → UMAP pipeline at the phase-ordering task itself: that gap (0.897 against 0.836) is smaller than the run-to-run variation above. These are single-dataset results measured against distributional published claims, which is a real limit on how far they generalize.

The non-determinism was root-caused rather than merely observed — an unseeded ARPACK start, a dropped seed with misaligned positional arguments in spectral_layout, a numba prange race in the MAP optimiser, and the seed never reaching PaCMAP/UMAP/TriMAP. Each fix is verified, and an upstream report is drafted but unsent.

Packaging findings

Hardening surfaced several upstream facts that affect anyone installing this stack: a git install reports version 1.0.2 for code identifying itself as 1.1.0 (the version bump was never committed, while the PyPI 1.1.0 sdist is correct); install_requires omits eight packages the paper-facing workflow needs, one of which is an unconditional import, so a bare install fails at import topo; conda-forge carries only 0.2.1.1, predating the published API; and scVelo 0.3.4’s default velocity mode cannot run under NumPy ≥ 2. The fixtures topometry-environment, topometry-1.1-environment, scvelo-environment and scanpy-environment exist because of these.

Redistribution

Restricted, and this is what blocks the study. The pancreas object is served from a GitHub repository with no license file, making it all-rights-reserved by default. It may be downloaded and analysed locally; it may not be vendored, subsampled and committed, containerized, or shipped as a test fixture. The tracked figures and numerical reports were produced from it, so the repository’s MIT code license does not extend to them and their status is explicitly unresolved.

An earlier assumption that a CC BY 4.0 Figshare fixture was available holds for some scVelo datasets but not this one.

What would make this complete

A redistributable fixture, which needs one of three things: an upstream license, regeneration from GEO GSE132188, or a synthetic ground-truth fixture. This is why the status here is blocked rather than running — the missing piece is a licensing resolution, not queued work.

The persistent-homology extension that would measure the cell-cycle loop quantitatively rather than by layout inspection has not started, and its dependencies are deliberately kept outside the replication environment so that a Foundry extension can never be mistaken for replication evidence.

topometry reviews the software. spectral-geometry covers the technique; the unstarted extension would use persistent-homology. hiponet-melanoma is the other study here where the released procedure’s evaluation protocol turned out to be part of the finding.