TopoQA: A Topological Deep Learning-Based Approach for Protein Complex Structure Interface Quality Assessment

Artificial Intelligence Information Science Bioengineering Biophysics Computer Science Biochemistry Life Sciences
Quality Assessment Protein Interface Persistent Homology Graph Neural Network Topological Data

Academic Background

The elucidation of protein complex 3D structures is a central topic in modern structural biology, molecular mechanism studies, drug design, and even artificial protein design. The function of a protein is often determined by its structure, and many biological processes involve complex interactions between proteins. Although traditional experimental methods (such as X-ray crystallography, cryo-EM, and NMR) can resolve protein 3D structures, they are time-consuming, costly, and less suitable for high-throughput or large-scale studies. In recent years, data-driven protein structure prediction methods (such as AlphaFold, RoseTTAFold, etc.) have achieved revolutionary breakthroughs, especially reaching experimental-level accuracy for monomeric proteins. However, the accuracy of protein complex structure prediction still lags behind monomer prediction, particularly in complicated systems such as multimers and antibody-antigen complexes, where much room for improvement remains.

In practice, researchers often need to select the most precise model closest to the native structure from a large number of “decoy” candidates generated by machine learning or deep learning models. In such scenarios, the key challenge is how to accurately evaluate and rank the quality of these candidate protein complex structures—namely, the so-called “Estimation of Model Accuracy” (EMA) or “Quality Assessment” (QA)—when the true structure is unknown. CAPS (Critical Assessment of Structure Prediction) experiments, as the authoritative competition in the field of protein structure prediction, have regarded EMA as an indispensable core part of the structure prediction workflow.

Traditional EMA/QA methods can be roughly divided into three categories: consensus-based, pseudo-single model, and single-model approaches. Consensus methods rely on the similarity between structures in the model pool; pseudo-single model methods generate their own pools for comparison. Both are limited by the quality of the pools and are computationally expensive. In comparison, single-model methods directly judge individual structural features, which are typically further divided into physics/statistical-potential methods and deep learning methods. In recent years, single-model QA based on Graph Neural Networks (GNNs) has demonstrated strong potential for feature extraction and global correlation capture but still struggles to describe high-order topological structure of protein complex interfaces.

To overcome these limitations, emerging algorithms such as Topological Data Analysis (TDA) and Persistent Homology (PH) have gradually been introduced into life science research. PH can detect and quantify high-order topological invariants (e.g., connected components, loops, voids) across multiple scales in complex systems, revealing novel structural features inaccessible to conventional graph models. Integrating GNNs with PH offers the promise of capturing more comprehensive information and achieving better generalization in protein complex quality assessment.

Source and Authors

This paper was co-authored by Bingqing Han, Yipeng Zhang, Longlong Li, Xinqi Gong (corresponding author, Renmin University of China), and Kelin Xia (corresponding author, Nanyang Technological University). It was published in Briefings in Bioinformatics (Volume 26, Issue 2, 2025, Article bbaf083) by Oxford University Press as an open access paper, released in 2025.

The team spans Renmin University of China and Nanyang Technological University, Singapore, with a strong background in mathematical sciences and structural bioinformatics, focusing on protein structure prediction, topological data analysis, and graph neural networks.

Detailed Research Workflow

1. Problem and Aim

The best protein complex structure prediction tools to date include AlphaFold-Multimer (AF-Multimer) and AlphaFold3 (AF3), yet significant shortcomings remain in interface accuracy assessment. While existing single-model GNNs can capture overall structural information, they tend to overlook high-order topological features at the atomic level, especially at protein complex interfaces. Therefore, the goal of this work is: to develop a novel protein complex interface quality assessment method, TopoQA, by combining Persistent Homology (PH) and GNN, achieving a deep integration of high-order topology and global coupling.

2. Dataset Construction

a. Training and Validation Data

  • Multimer-AF2 Dataset (MAF2): Contains complex structures predicted by AlphaFold2 and AF-Multimer, with target proteins from the EvCoupling and DeepHomo databases, totaling 9251 decoy models.
  • Dockground Dataset: 58 protein complex targets, each with an average of about 9.83 correct and 98.5 incorrect decoys.
  • Dataset Splitting: After sequence clustering at 30% homology, 70% of the data are used for training and 30% for validation. The final training set contains 8733 structures; the validation set, 3407.

b. Test Sets

  • DBM55-AF2: 15 antibody-antigen targets, 449 decoys.
  • HAF2: 13 heterodimer targets, 1370 decoys.
  • ABAG-AF3: 35 novel antibody-antigen targets, each generated by AF3 with 25 conformations per target, repeated five times (different random seeds), for a total of 875 structures.
  • All test sets undergo 30% sequence identity filtering to strictly prevent high-similarity data leakage.

3. Reference and Evaluation Metric System

  • Reference Metrics:

    • DockQ (integrates L-RMSD, I-RMSD, and Fnat for interface similarity; higher values indicate more accurate interfaces)
    • CAPRI, DockQ-wave, QS-score, etc.

  • Statistical Metrics:

    • Ranking Loss (measures the ability to select the best model accurately)
    • Top-10 Hits Rate (number of high-quality models among the top 10)
    • Pearson and Spearman correlation coefficients (linear/monotonic correlation between predicted and true scores)

4. Innovations in Topological Features and Graph Model Implementation

a. Persistent Homology (PH) Introduction

  • Essence and Workflow: Spatial point clouds of atoms surrounding a residue are divided by element (C, N, O, and their combinations), with 0-D PH (connected components) calculated via Vietoris-Rips complexes, and 1-D PH (loops) via Alpha complexes. The process is fully unsupervised, revealing local topological structure.
  • Barcode Vectorization: For each point cloud’s PH barcode, the birth and death times are used to extract mean, standard deviation, extreme values, and sum statistics. Each residue receives a 140-dimensional high-order topological feature vector.

b. Protein Interface Graph Representation

  • Graph Nodes: Residues within 10Å of the interface are selected as nodes.
  • Edge Feature Design: Besides Cα–Cα distance, the design innovatively introduces all-atom pair distance distributions between two residues. Distances are divided into ten bins, and the count in each is used to fill 10 edge dimensions; together with the Cα–Cα distance, this totals 11 edge features.
  • Basic Node Features: Amino acid type, secondary structure, relative solvent accessibility, dihedral angles, etc., for a total of 32 dimensions.

c. Proteinat: Custom GNN Module

  • Multi-head Attention: Node and edge embeddings are iteratively updated through attention mechanisms. The influence weights between nodes are determined by both node and edge features, with trainable optimization.
  • Global Feature Fusion for Prediction: Node and edge embeddings are pooled into graph-level features, passed into a multi-layer perceptron (MLP) to regress DockQ or similar scores. Mean squared error loss is used during training.

d. Overall Framework Workflow

Structure → Interface point cloud → PH topological encoding → Graph modeling and feature assembly → Multi-head attention GNN message passing → Global embedding → Predict interface DockQ score/rank

5. Comparisons and Baseline Models

  • Mainstream single-model deep learning approaches: GNN-DOVE, DProQA, ComplexQA, TRScore, etc., were chosen.
  • AF2Rank from AlphaFold-Multimer self-assessment module, and the latest AlphaFold3 IPTM (Interface Predicted TM-score), were also included for comparison.

Main Experimental Results

1. Comprehensive Benchmarking on Three Test Sets

a. DBM55-AF2 Results

  • TopoQA’s average Ranking Loss is 0.069, significantly lower than ComplexQA (0.26) and AF2Rank (0.261), representing reductions of 73.5% and 73.6%, respectively.
  • Among 15 targets, TopoQA directly selected the top structure for 4etq, 5y9j, and 6al0, with a Ranking Loss of 0.
  • Top-10 high-quality hit rates and correlation coefficients all outperformed other methods, demonstrating optimal and stable performance.

b. HAF2 Results

  • TopoQA’s average Ranking Loss is 0.11, superior to AF2Rank (0.12) and DProQA (0.192).
  • Top-10 hit rates for medium, good, and high-quality interfaces are all leading, with overall robust performance.

c. ABAG-AF3 Results

  • Apart from AF3, TopoQA achieved the lowest Ranking Loss (0.092), higher than DProQA (0.124) and ComplexQA (0.106).
  • For nearly half of the 35 tasks, TopoQA outperformed AF3’s main module, indicating the strong generalization and complementarity of the topology-based approach for novel structures.

2. Consistency across Multiple Reference Metrics

  • Under three evaluation systems—DockQ, QS-Score, and DockQ-Wave—TopoQA consistently ranked at the top for both Ranking Loss and correlation statistics, demonstrating high robustness and generalizability.

3. Ablation Study

  • Removing node PH topological features significantly weakened performance (Ranking Loss increased by up to 87%, correlation indices dropped 20–80%), highlighting the critical role of PH high-order topology in accurate structural assessment.
  • Removing atomic distance distribution edge features also decreased performance, indicating that fine-grained atomic-level information supplements traditional residue-centric properties.

Conclusion and Academic Value

TopoQA, as the first protein complex interface QA tool to integrate persistent homology-based topological features and GNN deep learning, not only surpasses state-of-the-art models (like AF2Rank, DProQA, ComplexQA) on major benchmarks, but also shows unique advantages on new and complex systems. Through topological encoding to capture high-order invariants in structures, its micro-level resolution of interface accuracy is greatly improved.

Scientific Significance: - Broadens protein structure representation from traditional sequence/physico-chemical attributes to high-dimensional topological invariants, advancing protein structural bioinformatics methodology. - Reveals the biophysical significance of high-order topological features (like connected components, loops) in distinguishing protein interfaces, providing new tools and insights for studying related molecular mechanisms. - Establishes a topological deep learning paradigm, offering a general reference for tasks such as molecular recognition, interaction prediction, and binding mode classification.

Application Value: - Can be directly integrated into mainstream workflows such as AlphaFold, AF-Multimer, AF3, improving model selection and accuracy assessment efficiency. - Applicable to active research areas involving structure screening, such as ligand binding, drug screening, protein design, enabling automated, large-scale, and high-throughput evaluation.

Research Highlights and Innovations

  1. First Large-scale Application of PH High-order Topological Features: Introduces PH high-order topology for protein interface assessment, greatly improving microstructure resolution.
  2. Introduction of Multi-scale Atomic Statistical Edge Features: High-dimensional histogram encoding of atomic distributions in edge features, enhancing local geometric structure representation at interfaces.
  3. Custom GNN Module Proteinat Design: Multi-head attention decouples node and edge deep learning, finely restoring information propagation paths in protein interfaces.
  4. End-to-end Extensible Framework: TopoQA can be seamlessly integrated into other protein AI prediction and screening systems, with broad compatibility.

Other Valuable Information

  • Open Data and Code Sharing: The authors have made the abag-AF3 dataset, TopoQA source code, and models publicly available (http://mialab.ruc.edu.cn/topoqa-master/code), facilitating community adoption and reproducibility.
  • Future Outlook: The team plans to further explore the integration of PH and deep learning, incorporate protein language models and other feature levels, promote multi-task learning, and enable comprehensive assessment ranging from local interfaces to global folding and local accuracy.

Summary

Accurate evaluation of protein complex structures is a critical foundation for the advancement of structural biology and molecular medicine. The innovation of TopoQA lies in the first combination of high-order topology and deep learning for protein interface assessment, providing an unprecedented new perspective for precise discrimination of protein interfaces. Its excellent performance across multiple authoritative datasets validates its technological advancement and practicality. As structure prediction AI enters the “complex era,” this method is set to play a pivotal role in broader structural biology research and practical applications, becoming a crucial driving force for precision medicine, drug design, and understanding molecular biological mechanisms.