HSSPPI: Hierarchical and Spatial-Sequential Modeling for PPIs Prediction

Artificial Intelligence Information Science Bioengineering Computer Science Biochemistry Genetics Life Sciences
protein–protein interaction sites hierarchical graph spatial form sequential form feature fusion

Background: Unveiling the Bottlenecks and Opportunities in Protein Interaction Prediction

Proteins serve as the core molecules for life activities, participating in almost all biological processes and cellular functions, including gene expression, RNA transcription, DNA synthesis, immune response, and more. Protein-protein interactions (PPI), as well as interactions at specific sites (Protein-Protein Interaction Sites, PPIS), determine the diversity and precision of physiological activities. For example, drug design, protein function annotation, exploration of disease molecular mechanisms, and the construction of global protein interaction networks all rely on high-quality PPI and PPIS information.

However, traditional experimental approaches—such as X-ray crystallography and mass spectrometry—for detecting PPI sites are time-consuming, costly, and face issues like high sample complexity and limited scalability. With the rapid expansion of protein databases and urgent needs for disease prevention and control, computational methods for PPIS prediction have emerged vigorously. These current methods mainly fall into two categories: sequence-based and structure-based, most of which rely on machine learning or deep learning models to mine potential protein interaction information.

Although these methods have achieved some progress, there remain prominent limitations: (1) Most methods only focus on a single representation form of proteins—either spatial conformation or primary sequence—failing to integrate both sequence and structural information. This shortcoming prevents simultaneous insight into the temporal and spatial proximity of neighbor residues, thus limiting predictive performance; (2) Many existing models neglect the intrinsic hierarchical structure of protein molecules. For example, a protein’s residues are composed of atoms, and functional binding often requires attention to underlying atomic-level details. Traditional methods extract features mostly at the residue level, rarely capturing complex atomic-level information such as amino acid side chains with precision, thus hampering expression power.

Thus, deeply integrating the hierarchical structure and both spatial-sequential information of proteins has become the key challenge to boost PPIS prediction.

Paper Source and Author Introduction

The research paper entitled “hssppi: hierarchical and spatial-sequential modeling for ppis prediction” was completed by Yuguang Li, Zhen Tian, Xiaofei Nan, Shoutao Zhang, Qinglei Zhou, and Shuai Lu. The authors come from multiple research institutions including the School of Computer and Artificial Intelligence at Zhengzhou University, Yangtze Delta Region Institute (Quzhou) of the University of Electronic Science and Technology of China, School of Life Sciences (Zhengzhou University), Zhongyuan Intelligent Medical Laboratory, and the National Supercomputing Center in Zhengzhou, among others. The paper was published in 2025 in the internationally renowned journal Briefings in Bioinformatics by Oxford University Press.

Research Workflow and Analysis of Innovative Techniques

Overall Research Architecture

In response to the need for accurate prediction of protein-protein interaction sites, this study proposes a brand-new deep learning framework—HSSPPI (Hierarchical and Spatial-Sequential Protein-Protein Interaction predictor). For the first time, this model represents proteins as complete hierarchical graph structures and jointly models their spatial arrangements and sequential order, while innovatively introducing spatial-sequential (s-s) modules to achieve multi-dimensional, multi-scale, deep integration of intrinsic protein information.

The overall architecture of HSSPPI encompasses four modules: hierarchical protein graph generation, atom-level feature extraction, residue-level feature extraction, and final feature classification and prediction.

Step One: Hierarchical Protein Graph Generation

  • Hierarchical Representation: The authors build the protein as a “residue-atom” nested dual-level structure. At the macroscopic level, protein nodes are residues, each further refined into an atom-level graph (atom-level graph).
  • Connection Relations: Connection boundaries are defined using distance thresholds between atoms or residues. The presence of an edge is determined by measuring the Euclidean distance between key heavy atoms.
  • Parameter Exploration: Atom-level edges are assigned candidate thresholds from 1.3 to 2.5 Å; residue-level edges use a range of 4.0 to 7.0 Å. This lays the foundation for subsequent fusion experiments.

Step Two: Atom-Level Feature Extraction

  • Feature Encoding: All 37 atom types are one-hot encoded, resulting in an atom feature matrix.
  • Spatial-Sequential Block (s-s block): The key s-s block, composed of a Graph Convolutional Network (GCN) and a Bidirectional Gated Recurrent Unit (Bi-GRU), is introduced to aggregate spatial neighbors and contextual sequential information:
    • The GCN captures spatial correlations and neighborhood topology among atoms;
    • The Bi-GRU simulates the intrinsic semantics and local grammar of protein sequences—enabling both forward and backward information flow.

Step Three: Residue-Level Feature Extraction

  • Feature Acquisition: The pretrained protein language model ProtT5 is employed to produce a high-dimensional embedding vector (length 1024) for each residue.
  • Adjacency Matrix Construction: The adjacency matrix is built by calculating the mean atomic distance between residues.
  • Spatial-Sequential Fusion & Hierarchical Integration: As in the atom-level, GCN and Bi-GRU are linked in tandem. Feature-Fusion blocks (f-f block) are designed to concatenate atom-level and residue-level features according to subordination, thoroughly coupling multi-level information.

Step Four: Feature Classification and Prediction

  • Skip Connections & Dual Fully Connected Layers: The final fused features are processed via skip connections and two fully connected layers for classification, thereby predicting the interaction site for every residue.
  • Activation & Loss Function: ReLU activation and weighted cross-entropy loss are used to optimize recognition accuracy.

Algorithm Implementation

  • Hardware/Software Environment: Implemented with PyTorch, training is performed on a single NVIDIA RTX4090 GPU, with each epoch taking roughly 20-30 minutes.
  • Hyperparameter Search: Learning rate, dropout rate, and distance thresholds are comprehensively optimized using cross-validation.
  • Dataset Setup: Large-scale comparison experiments are conducted on multiple public benchmark datasets from two authoritative PPIS tasks—DeepPPISP and GraphPPIS—with total sample counts in the thousands, and each containing training, test, and independent validation sets.

Results and Findings

DeepPPISP Task Comparison

  • Comparison Group: Eight mainstream models—Sppider, Scriber, DeepPPISP, Attention-CNN, Delphi, HN-PPIS, EGRET, EnsemPPIS—are benchmarked alongside.
  • Accuracy Superiority: HSSPPI stands out particularly on imbalanced data. Average improvements in F1 score, MCC (Matthews Correlation Coefficient), and AUPRC (Average Precision) all significantly surpass baseline models. F1 increases by 50.23% over the best prior model, MCC by 15.16%, and AUPRC by 7.16%.
  • Generalization Performance: Not only are the overall scores high, but the precision index increases by 68% compared to EnsemPPIS, and recall by 30.83%, indicating the model greatly reduces false positives without losing highly relevant sites.

GraphPPIS Task Comparison

  • Expanded Group: In addition to the above, a total of 11 more methods are included, such as PSIVER, PRONA2020, DLPred, MaSIF-Site, GraphPPIS, RGN, Prob-Site, AGAT-PPIS, DeepProSite, GHGPR-PPIS, and GACT-PPIS.
  • Multi-Metric Lead: Whether compared against structural or sequence-based models, HSSPPI sets records for Precision, Recall, and F1 across key metrics. Notably, compared to GACT-PPIS, these three see respective improvements of 9.9%, 1.1%, and 7%.
  • Summary: Although it trails in some metrics like Acc, HSSPPI achieves the best performance on important metrics sensitive to imbalance (AUPRC, AUC, and F1), aligning with trends and real-world needs in the post-AlphaFold era.

Generalization and Robustness

  • Independent Testing: Across three independent test sets—Test287, TestB25, and TestUB25—HSSPPI secures first place or prominent improvements in F1 and other core metrics, demonstrating strong adaptability and generalization, as well as sensitivity to protein conformational changes.

Ablation Studies and Model Selection

  • Hierarchical Ablation Comparison: Using only the atom-level or only the residue-level graph both underperform compared to their fusion, confirming that multi-level feature fusion underpins accurate prediction. Over-fusing leads to over-smoothing; moderate fusion (twice) is optimal.
  • Threshold Optimization Experiments: Combined experiments reveal the best AUROC and AUPRC are achieved when the residue threshold is 5.5 Å and the atom threshold 2.3 Å, validating the effectiveness of the spatial adjacency strategy.
  • GCN Architecture Comparison: Multiple graph neural networks are tested—GraphConv, SageConv, ChebConv, GATConv, GATV2Conv, GCNConv—ultimately selecting SageConv as the core convolution operator for its balance of accuracy and efficiency.

Visualization and Case Analysis

  • Case Validation: Taking the protein with PDB ID 1B6C_A as an example, the concordance between actual interaction site annotation and HSSPPI’s prediction exceeds competing methods. Even after the protein’s conformation is rotated, key binding sites are still accurately located, with the MCC value showing a 15.5% increase over EnsemPPIS.

Research Conclusions and Value

The HSSPPI model successfully overcomes the bottlenecks of existing methods in protein-protein interaction site prediction, achieving the following scientific and practical value:

  • Scientific Contribution
    • First to describe proteins as hierarchical graph structures, systematically fusing atomic- and residue-level information to enhance protein representation.
    • Innovatively designed a spatial-sequential joint extraction module allowing temporal and spatial neighbor information to be aggregated in parallel, fully revealing the structure-function relationship.
    • Enables large-scale parallel training and generalizes well, suitable for complex, imbalanced task scenarios.
  • Applied Significance
    • Provides higher-precision, more interpretable solutions for new drug discovery, protein function annotation, and disease molecular mechanism analysis.
    • The model is “plug-and-play”, highly compatible and seamlessly connectable with the latest structural prediction trends such as AlphaFold-Multimer and RoseTTAFold All-Atom.
    • Open-source code enhances reproducibility and extensibility in the field.
  • Methodological Highlights and Unique Issues
    • The unique hierarchical graph structure and feature fusion strategy set an example for protein structure integration and cross-scale modeling;
    • The S-S Block links GCN and Bi-GRU, implementing spatial and sequential neural network processing in protein structure modeling for the first time;
    • Fine-grained comparison across multiple datasets, scenarios, and architectures with solid ablation studies, making the conclusions highly reliable.

Other Valuable Information

  • Funding Support: The research is funded by the National Natural Science Foundation of China, Henan Provincial Natural Science Foundation, and related major science and technology projects.
  • Code Availability: The HSSPPI project code is hosted on GitHub, facilitating sharing and application within both industry and academia.
  • Outlook: The team plans to further incorporate “interaction partner” features into feature extraction, pushing the precision limit for protein-protein complex prediction.

Conclusion

This research offers profound insight into the essential properties of protein molecules and the bottlenecks of bioinformatics mining, fusing cutting-edge deep learning and structural biology. The new methodological system established here achieves higher accuracy and stronger generalization for protein-protein interaction prediction, marking a breakthrough for intelligent proteomic analysis and delivering a valuable methodological template for future algorithm development in related fields.