TriGraphQA

Two residues in contact inside a single chain and two residues in contact across a binding interface are physically different events: one reports whether a chain has folded correctly, the other whether two chains have been docked correctly. Graph-based quality assessment usually flattens both into one homogeneous residue graph, where the two kinds of edge are the same object and the network must rediscover the distinction from features alone.

TriGraphQA keeps them apart by construction. A dimer is represented as three graphs: a residue-node graph for each chain, capturing intra-chain folding environment, and a contact-node graph for the interface, in which the nodes are inter-chain residue pairs and an edge joins two contacts that share a residue. The two monomer graphs are encoded first, and an interface context aggregation module then projects those embeddings onto the interface — for a contact between residue i in chain A and residue j in chain B, every other residue in chain A contributes to i with a weight proportional to the log of its sequence separation divided by its Cα distance. That weighting favours residues far apart in sequence but close in space, the long-range tertiary contacts that indicate a chain is genuinely folded, so an interface node learns how well the monomer around it supports it.

Luozhan Liang and Kailong Zhao at Shandong University posted the work to bioRxiv in March 2026 and published it in the Journal of Chemical Information and Modeling in August 2026. It joins the complex quality assessment line the catalog already carries — TopoQA, ContrastQA, GraphCPLMQA and DeepUMQA3 — and makes graph decomposition itself the contribution.

#Key Features

• Triple-graph decomposition: Intra-chain folding stability and inter-chain binding specificity are learned in separate graphs, not inferred from one shared representation.
• Contact-node interface graph: Nodes are inter-chain residue pairs within 7 Å heavy-atom distance and edges connect contacts sharing a residue, so message passing runs directly on the geometry of the binding patch.
• Interface context aggregation: Sequence-distance-aware weighting gives a 64-D context vector per participating residue, concatenated with a 128-D projection of the contact's own features into a 256-D embedding.
• Gated graph convolution: Edge features modulate every message through a sigmoid gate, three layers to a block.
• Checkpoint-only inference: A feature-build step converts a folder of PDB complexes into graphs and a predict step scores them against the shipped checkpoint, with no per-target fitting.

#Technical Details

Residue nodes carry 41 features — a 20-D amino-acid one-hot, 7 Meiler descriptors, backbone bond geometry, Rosetta one-body energies and secondary structure — with 21-D edges; contact nodes carry 37, including Rosetta two-body terms and a VoroIF-GNN-inspired orientation encoding, with 26-D edges from radial basis functions and angular bins. Training data came from Q-BioLip (January 2024 cutoff), filtered to 10,502 dimers with monomers under 500 residues; decoys were generated with AlphaFold 3 across 25 random seeds and 30 HDOCK poses per target, then sampled into four DockQ quality bins to give 168,061 structures, with 40% sequence identity to the test sets removed to leave 10,352 training targets. On the held-out Dimer50 benchmark it reaches Pearson 0.534 and Spearman 0.417 against true DockQ, with a top-1 ranking loss of 0.174 and mean top-1 DockQ of 0.646, above MViewEMA (0.294 loss), TopoQA (0.298) and ComplexQA (0.414). On DBM55-AF2 it records a 0.128 top-1 loss and 0.487 mean top-1 DockQ; on HAF2 it reaches AUC 0.920 and 0.945 for separating acceptable and medium-quality decoys, though TopoQA's 0.110 top-1 loss there beats its 0.134. Ablations replacing the triple graph with a global-plus-interface pair (0.288 loss) or an interface-only graph (0.335) confirm the decomposition, not the feature set, carries the gain.

#Applications

The use case is model selection: given the pool of poses a docking run or an AlphaFold-Multimer job returns for one target, TriGraphQA ranks them so interface analysis, mutational design or refinement starts from the best available structure. Its benchmarks emphasise heterodimers and antibody-antigen targets, where folding-model confidence is least reliable, and the feature pipeline needs Rosetta and a Voronota binary in place before scoring.

#Impact

TriGraphQA's result is architectural: the ablation shows that fusing monomer and interface information at the feature level beats both a single homogeneous graph and a late average of two separate scores. The framework is built for dimers, and multimeric assemblies are handled by decomposing them into pairwise dimers and averaging — using the native structure to decide which pairs genuinely interact, a step available at benchmark time but not prospectively, which the authors argue acts as a constant factor within a target and so does not disturb ranking. Higher-order interactions are named as future work. The repository carries inference code and a 5 MB pretrained checkpoint but no training code and no license file, leaving reuse terms undefined, and the released Dimer50 benchmark is a single archive with no accompanying documentation.