ContrastQA

Modern predictors emit many candidate structures per complex and no reliable ranking among them, which makes Estimation of Model Accuracy (EMA) — scoring a predicted structure without knowing the answer — a load-bearing part of any structural pipeline. EMA for single-chain proteins is mature; EMA for complexes is not. The difficulty is that complex quality is dominated by the interface, a small fraction of the structure, while the score being predicted is global. A network tuned to interface geometry can miss a globally wrong assembly, and one tuned to global shape can be fooled by a plausible fold with a misplaced partner.

ContrastQA, developed by Renzhi Cao's group at Pacific Lutheran University with collaborators at Anhui University, Hamilton College, Saint Louis University, and University of Washington Bothell, resolves this by supervising a contrastive representation with interface quality. Its key idea is label-guided contrastive learning: rather than constructing positive and negative pairs by augmentation, ContrastQA selects them using DockQ interface scores, and weights each negative by how far its DockQ value sits from the anchor's. The resulting embedding space is organized by interface quality, and a geometric graph network reads global structural context on top of it.

Posted to bioRxiv in June 2025, the model was evaluated zero-shot on CASP16 targets it never saw during training, where it reduces ranking loss by 8.7-14.9% relative to the best competing EMA method depending on the metric and target difficulty.

#Key Features

• Label-guided pair selection: Positives and negatives come from DockQ interface scores rather than data augmentation, so the contrastive signal reflects genuine quality differences instead of synthetic perturbations.
• Quality-weighted contrastive loss: A modified NT-Xent objective weights same-target negatives by the absolute DockQ gap to the anchor, making near-indistinguishable decoys contribute proportionally less than clear failures.
• Interface and global information combined: A geometric vector perceptron graph network models whole-structure geometry while the contrastive objective supplies interface-level discrimination.
• Frozen protein language model features: Per-residue embeddings from the final layer of ESM-2 enrich each node with evolutionary context that geometry alone does not carry.
• Transferable pretraining strategy: The authors frame the contrastive module as a general pretraining recipe for protein structure representations, not solely an EMA component.

#Technical Details

Each candidate complex becomes a k-nearest-neighbor graph over residues with k = 10. Node features combine 55 geometric and physicochemical dimensions with 1280-dimensional ESM-2 layer-33 embeddings, fused by an MLP into 512 dimensions; edges carry 22 features. A GVP-GNN provides rotation-equivariant message passing, and the contrastive head uses a temperature of about 0.05 with up to eight negatives per anchor.

Training pooled roughly 20,000 decoys: 4,000 from the MULTICOM set (40 targets), 1,750 generated with AlphaFold-Multimer across 70 targets, 8,860 generated with AlphaFold 3 across 443 targets, and 10,126 from the VoroIF-GNN heterodimer set, split roughly 8:2 with sequence identity between training and validation held below 30%. On CASP16 (37 targets, 12,142 decoys) ContrastQA records top-1 ranking losses of 0.123 on TMscore and 0.116 on GDT-TS, against 0.138 and 0.128 for the next-best method; on CASP16 hard targets it reaches 0.131 versus 0.154. On an antibody-antigen set built with AlphaFold 3, it records 0.028 TMscore loss against 0.031 for VoroIF-GNN and 0.037 for TopoQA. Comparisons also include GNN-DOVE, ComplexQA, and DProQA.

#Applications

The intended use is model selection: given a stack of predicted complexes for a target, ContrastQA ranks them so downstream work proceeds from the best available structure. That matters most where the assembly, not the fold, is the scientific question — antibody-antigen complexes, receptor-ligand pairs, and multi-subunit machines — and it complements monomer-oriented quality estimators such as ELEN, which score local regions of single chains rather than assemblies.

#Impact

ContrastQA's contribution is methodological: it shows that a continuous structural quality label can supervise contrastive pair construction directly, which is a transferable idea beyond complex EMA. Its limitations are stated plainly by the authors. The ~20,000-decoy training set is two orders of magnitude smaller than what leading CASP16 EMA groups used; DockQ alone is an imperfect basis for partitioning pairs; hardware limits capped negatives at eight per anchor; and the method targets global accuracy only, leaving interface-level and per-residue EMA out of scope. The work is a preprint, and while training code and datasets are on GitHub the repository carries no license file, which leaves reuse terms undefined.