Loading…
Protein complex interface quality assessment model that pairs persistent homology barcodes with a graph attention network to predict DockQ scores.
No providers recorded yet. Browse all providers
A structure predictor hands back twenty-five candidate assemblies for the same complex and no trustworthy way to tell which one is right. Quality assessment (QA), or estimation of model accuracy, is the step that picks one — scoring a predicted complex without the native structure to compare it against. Deep-learning QA methods almost all represent the complex as a graph of residues joined by contacts, which captures who touches whom but flattens the shape of the atoms around each residue: two residues with identical contact patterns can sit in packing environments of very different geometry, and the graph cannot tell them apart.
TopoQA answers that with persistent homology. Around each interface residue it takes the neighbouring atoms as a point cloud, builds a filtration of simplicial complexes over it, and records when topological features — connected components, loops — are born and die. Those barcodes become a fixed-length fingerprint of the local atomic arrangement, attached to the residue node before any message passing happens, so the graph network works on nodes that already encode the higher-order structure of their surroundings.
The model came from Xinqi Gong's Mathematical Intelligence Application Laboratory at Renmin University of China, Kelin Xia's group at Nanyang Technological University, and collaborators at Shandong University, posted to arXiv in October 2024 and published in Briefings in Bioinformatics in March 2025. It joins the single-model QA line the catalog already carries — DeepUMQA, GraphCPLMQA, ContrastQA — as the first to build its node representation from topological data analysis rather than sequence embeddings or voxelized environments.
Each node carries 172 features: the 140 topological dimensions plus 32 basic ones — a 21-dimensional residue-type one-hot, an 8-dimensional secondary-structure one-hot, relative solvent accessible surface area, and two torsion angles. Training used 8,733 conformations spanning dimers through heptamers, drawn from an AlphaFold2 and AlphaFold-Multimer decoy set and the Dockground resource, under the same splits as DProQA and ComplexQA and a mean-squared-error objective against DockQ.
On DBM55-AF2 (15 antibody-antigen targets, 449 decoys) TopoQA reaches a ranking loss of 0.069, 73.6% below the 0.261 of AlphaFold-Multimer-based AF2Rank, though DProQA's 0.049 is lower. On HAF2 (13 heterodimer targets, 1,370 decoys) its 0.11 is the best of any method tested, 8.3% below AF2Rank and 42.7% below DProQA. On the authors' new ABAG-AF3 set (35 targets, 875 conformations generated by AlphaFold 3) it scores 0.092 against AF3's own confidence-based 0.054, while beating AF3 on 17 of the 35 targets. Ablating the topological features degrades the ranking loss from 0.069 to 0.129 on DBM55-AF2 and from 0.11 to 0.159 on HAF2.
The practical use is model selection: given the stack of complexes a predictor returns for one target, TopoQA ranks them so downstream work — interface analysis, mutational design, docking refinement — starts from the best available structure. Its benchmarks are weighted toward antibody-antigen and heterodimer targets, where folding-model confidence scores are least reliable. The network is small enough to score large decoy pools cheaply, though the feature pipeline needs DSSP, GUDHI and PyTorch Geometric in place first.
TopoQA's contribution is the demonstration that persistent homology features carry information a residue-contact graph does not: the ablation, not the leaderboard, is the result that matters, and it is large on both benchmark sets. The authors are explicit about what the model does not do. It scores interface accuracy only, leaving global fold accuracy and per-residue estimates to other methods, and complexes with no detected interface are filtered out rather than scored. Its 8,733-conformation training set is orders of magnitude smaller than the pools leading CASP15 QA teams used, which the authors name as the likely ceiling on performance. The released repository carries inference code and the pretrained checkpoint but no training code, and no license file, leaving reuse terms undefined.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.