open-topoqa-scorer

This fixture completes the open structure-QA path: the ProteinGAT scorer that consumes open-topoqa-featurizer-environment‘s graphs and predicts a DockQ-like interface-quality score. See topoqa for the upstream profile and the replication evidence.

The architecture is reproduced from the paper (Han 2025, Eqs 3-9). The paper pins no training or capacity hyperparameters, so the weights are our retrain on a reassembled MAF2 and Dockground corpus rather than the released checkpoint. As with the featurizer there is no bit-exact oracle, and divergence is correct: an independent reproduction confirmed the released code carries an (x, y, y) coordinate defect that roughly a third of the paper’s HAF2 ranking-loss margin depends on. Measured against a corrected (x, y, z) TopoQA, the retrain is at parity — it matches or beats every correlation and reaches HAF2 ranking-loss parity.

The environment stages two in-repo path recipes, the scorer and the featurizer it consumes, since the latter is on no public channel yet, and it drags a full pytorch and pytorch_geometric closure. The linux-64 lock solves cleanly (pytorch 2.13.0 cpu-mkl, pytorch_geometric 2.8.0, gudhi 3.13.0, dssp 4.6.1, biopython 1.87) and the noarch wheel builds and imports.