The closest recoverable released procedure reached 67.27% against a reported 90.90%, on a platform arm that cannot yet settle why.

HiPoNet melanoma replication

What was tested

hiponet predicts patient-level labels from high-dimensional cellular point clouds using multi-view simplicial complexes. The claim under test is its melanoma result: 90.90% patient-level classification accuracy on the Yale melanoma cohort.

The cohort was reconstructed from the exact Mendeley Data v1 URLs, verified by byte size and SHA-256 before use, and written as 54 ordered patient point clouds without splitting or subsampling.

What came out

The closest recoverable released procedure reached 67.27% mean best-test accuracy against the reported 90.90%. A controlled intervention using the paper’s wavelet formulation reached 69.09% mean accuracy and 61.39% balanced accuracy.

The outcome is recorded as inconclusive rather than as a failure to reproduce, and the distinction is not politeness. The completed arm ran on Apple MPS and is labeled a portability arm; the locked Linux/CUDA target that would separate a platform effect from a procedural one has not run. A number produced on the wrong platform cannot convict the method. What can be said is narrower and still substantive: the closest released procedure did not recover the headline in the environment that completed.

The hardening arm found something independent of that. The released procedure selects its best epoch on test patients. Under a corrected evaluation that reserves test patients until after selection, the best transparent arm is a cell-count diagnostic at 56.10% consensus balanced accuracy, with a 95% bootstrap interval of 44.18–68.02% — an interval containing chance.

The extension arm asked whether explicit topological summaries predict response at all: H0/H1 summaries reached 49.72% and persistence images 45.93% consensus balanced accuracy. No arm robustly clears chance.

That last result is easy to over-read. It does not establish that no biological signal exists. It establishes that the recovered cohort, under prespecified transparent representations and a leakage-safe protocol, does not support a robust patient-level signal. Those are different claims, and the study keeps its tracks separate precisely so they are not mixed.

An upstream hazard this corpus has hit before

Building boundary matrices from a simplex tree can silently drop isolated vertices — invisible on connected inputs, fatal on bipartite constructions. It is recorded under simplicial-learning and is worth checking in any higher-order pipeline of this shape.

Redistribution

Mixed, and structurally so. The core — data retrieval, cohort reconstruction, checked contracts, transparent baselines, metrics, and evidence verification — is MIT. The faithful and hardened neural runners derive from upstream HiPoNet and remain under Yale’s non-commercial license, so they live in a physically separate package that importing the core does not pull in. That separation is the reason the MIT half is usable at all. Raw data, prepared arrays, run trees, and checkpoints are deliberately not in version control.

What would make this complete

Two things, in order: the locked Linux/CUDA arm, which is execution-ready at 25 folds and would settle whether the accuracy gap is platform or procedure; and then a biopixi fixture re-running the pinned repository here. The non-commercial license ceiling on the HiPoNet-derived component constrains what such a fixture could distribute, which is the same ceiling recorded in the hiponet note.

hiponet reviews the paper and software. simplicial-learning covers the technique; topological-deep-learning the broader practice. topometry-cell-cycle is the other study here where run-to-run variation turned out to be part of the finding.