open-topoqa-scorer

This is the model half of the open structure-QA path: a graph attention network that reads the interface graphs open-topoqa-featurizer emits and predicts a DockQ-like interface-quality score between zero and one. Ranking candidate structures for one target is what it is for, and it is a ranking aid rather than a statement that any candidate is biologically correct.

Why it was written

The same reason as the featurizer, one step further along. topoqa ships inference code and a pretrained checkpoint under no license, so neither could be redistributed. But a checkpoint has a second problem the featurizer did not: even if it were licensed, it was trained against the featurizer that builds edge coordinates as (x, y, y), so it is fitted to inputs a correct featurizer does not produce. Reusing those weights was neither permitted nor coherent, which is why the weights here are a retrain on a reassembled MAF2 and Dockground corpus rather than a port.

Divergence from the released checkpoint is consequently the correct outcome and not a shortfall. The comparison is recorded in topoqa: measured against a corrected (x, y, z) upstream, the retrain matches or beats every correlation and reaches ranking-loss parity on HAF2. topoqa-interface-quality holds the upstream-side evidence that comparison rests on.

What the paper pins, and what is ours

The architecture is the paper’s and is followed exactly: additive edge-conditioned multi-head attention over the source node, the target node, and the edge embedding; edges updated at every layer from the two updated endpoints and the previous edge; the updated edges mean-pooled and reduced to half the node width, concatenated with the pooled nodes, passed through three stacked linear layers and a sigmoid; mean-squared error against DockQ.

Everything else is ours, and the distinction is not a hedge. Neither the main text nor the supplementary tables state a value for the number of layers, attention heads, embedding widths, dropout, optimizer, learning rate, epochs, or batch size — the supplement’s tables are results and ablations, not hyperparameters. Those defaults are chosen from common practice and tuned on validation. They are never claimed as reproductions, and a reader comparing this model to the published one should treat capacity as an uncontrolled difference.

One departure is deliberate rather than forced. Mean-squared error alone lets the model collapse a decoy set onto its target mean: absolute fit looks reasonable while within-target ordering approaches random, which is the one thing a selection model must not do. An optional within-target pairwise hinge penalizes mis-ordering two decoys of the same target and can be added to the loss. It is off by default, so the default training objective stays the paper’s.

What it is worth against the alternatives

Two results in the repository’s benchmark record are worth carrying into any decision to use it. Against a trained non-topological competitor and against training-free AlphaFold-confidence baselines, run on identical decoys under one metric implementation, the model’s advantage appears clearly in rank correlation and top-N success — and largely disappears in top-1 ranking loss, where mean interface pLDDT is about as good. Reporting only the metric that favors the method would misrepresent it.

The second is that the residual gap to upstream localizes to a couple of targets whose single top-ranked decoy is mis-selected, and survived a capacity, dropout, epoch-budget, and checkpoint-selection sweep unchanged. That points at an unspecified upstream training detail rather than at a reproducible component, and it is the honest boundary of what a paper-only reproduction reaches.

A packaging gap worth knowing about

The distributed package carries the model, the data adapter, the training loop, and the metrics. It does not carry weights: the trained checkpoints live in the repository, outside the wheel’s package directory. Anything installing this from a channel gets machinery that must be trained or pointed at a checkpoint obtained separately, which is not what a reader arriving from the recipe would assume.

In this corpus

open-topoqa-scorer-recipe builds it and open-topoqa-scorer-environment stages it together with the featurizer, since that dependency is on no public channel yet; both are authority for versions and dependencies. score-docking-poses is the procedure that puts the pair to work. The technique is topological-deep-learning and the subject it replaces is topoqa.