The released TopoQA checkpoint does reproduce the paper's ranking losses — and part of the margin depends on a coordinate defect.

TopoQA interface-quality replication

What was tested

topoqa ranks predicted protein-complex structures by interface quality, combining persistent-homology features with graph attention. The claim under test is its headline benchmark result: mean ranking loss on DBM55-AF2 and HAF2, plus the reported correlations.

The released checkpoint was run on the same benchmarks the paper reports, obtained from the DProQ distribution. Because the upstream repository declares no license, no TopoQA code, checkpoint, or benchmark structure is vendored into the replication repository — it orchestrates against a checkout the user supplies.

What came out

The released artifacts reproduce the paper. DBM55-AF2 mean ranking loss is 0.069400 against a reported 0.069; the paper-filtered HAF2-12 loss is 0.110333 against 0.110. The correlations reproduce too — but only as pooled statistics (Spearman/Pearson 0.5016/0.5147 on DBM55-AF2, 0.6747/0.6003 on HAF2-12). Mean per-target correlations are substantially lower. This is worth carrying forward whenever TopoQA is compared against a method reporting per-target numbers, because the two are not the same measurement and the paper does not label which it used.

Part of the reported margin depends on a coordinate defect. The released code builds all-atom edge-histogram coordinates as (x, y, y) rather than (x, y, z). Substituting z at inference time — retaining the released checkpoint — moves ranking loss to 0.076733 on DBM55-AF2 and 0.147083 on HAF2-12.

That second result needs its limits stated plainly. It is a sensitivity arm, not a corrected benchmark: changing coordinates at inference while keeping a checkpoint trained on the defective path measures how much the checkpoint depends on that path. It does not estimate what a correctly trained model would score. Retraining was not available within this study.

Where this fits

The clean-room reimplementation is a separate line of work — open-topoqa-featurizer and open-topoqa-scorer, both MIT — and it reaches parity with a corrected upstream on HAF2-12. The comparison between the two is recorded in the topoqa note. To keep provenance clean, the upstream artifacts here were executed only by an independent third party on separate infrastructure, which returned per-decoy predictions and provenance facts rather than code. Every metric was then recomputed under one definition, so no difference in this note can be attributed to differing metric code.

Redistribution

Mixed, and the split matters. The replication code and its documentation are MIT. Upstream TopoQA declares no license and the Zenodo benchmark record identifies none, so neither is redistributed. The compact result tables derived from those inputs are marked NOASSERTION rather than MIT — a permissive license on the analysis code does not relicense outputs computed from unlicensed inputs.

What would make this complete

A biopixi fixture that re-runs the pinned repository inside this Foundry. Until that exists the evidence is real but lives only upstream, and this note pins it rather than reproduces it. The obstacle is not licensing here but the historical PyTorch stack the original model requires, which is deliberately outside the locked MIT verifier environment.

topoqa carries the full review of the paper and software. topological-deep-learning places the topology-plus-attention pattern in context; persistent-homology covers the featurization. hiponet-melanoma is the other study here whose headline number came out differently from the paper’s.