DeepPhosPPI: a deep learning framework with attention-CNN and transformer for predicting phosphorylation effects on protein-protein interactions

Yinyin Gong, Rui Li, Yan Liu, Jilong Wang, Danny Z. Chen, Chee Keong KwohBriefings in Bioinformatics 2025 · DOI

TL;DR

This paper introduces DeepPhosPPI, a novel deep learning framework that accurately predicts how protein phosphorylation affects protein-protein interactions (PPIs). By combining advanced neural network architectures with powerful protein language models, DeepPhosPPI offers a significant leap forward in understanding these crucial biological processes, with potential implications for disease research and drug development.

What Problem Does This Paper Address?

Protein phosphorylation, the addition of a phosphate group to a protein, is a fundamental mechanism that regulates a vast array of cellular functions, including protein-protein interactions (PPIs). These interactions are the building blocks of cellular signaling networks and are critical for maintaining cellular homeostasis. Dysregulation of phosphorylation can lead to serious diseases like cancer and neurodegenerative disorders. While experimental methods exist to study these effects, they are often costly, time-consuming, and not scalable for large-scale analysis. Existing computational methods have limitations: structure-based approaches require high-resolution 3D structures (which aren't always available) and are computationally intensive, while sequence-based methods often rely on manually engineered features and struggle to capture complex, long-range dependencies within protein sequences. This paper aims to bridge this gap by developing an efficient and accurate deep learning framework that leverages protein sequences to predict both functional phosphorylation sites and their specific regulatory effects on PPIs.

Methods

DeepPhosPPI is a two-stage deep learning framework:

  • Task 1: Functional Phosphorylation Site Identification (DeepPhosPPI-1): This stage identifies which serine (S), threonine (T), or tyrosine (Y) residues are likely to be phosphorylated and functionally relevant. It uses a convolutional neural network (CNN) with an attention mechanism. The framework starts by embedding protein sequences using pre-trained protein language models (PLMs) like ProtBERT or ESM-2, which capture rich semantic and biophysical features. These global embeddings are then processed by a hierarchical CNN structure to extract local features relevant to phosphorylation. An attention mechanism helps the model focus on the most important contextual information around potential phosphorylation sites.
  • Task 2: Regulatory Effect Classification for PPIs (DeepPhosPPI-2): This stage predicts whether a functional phosphorylation site enhances or inhibits the interaction between two proteins. DeepPhosPPI-2 employs an ensemble approach, combining two specialized models:
  • AttCNN-PhosPPI: This model uses a CNN with an attention mechanism to extract local features around phosphorylation sites and model their impact on PPIs.
  • Transformer-PhosPPI: This model utilizes a Transformer architecture to capture long-range, global dependencies within protein sequences, which are also crucial for understanding PPI regulation.

The predictions from AttCNN-PhosPPI and Transformer-PhosPPI are then fused using a "soft voting" ensemble strategy to produce the final classification of the regulatory effect (enhancement or inhibition). A novel "STY-Sliding Window" strategy is also introduced to specifically extract local features around S, T, and Y residues for the regulatory effect prediction task.

Datasets

The study utilizes two primary datasets:

  • Dataset A: Used for training and evaluating the functional phosphorylation site identification model (Task 1). It was curated from the PTMint database, containing experimentally validated functional phosphorylation sites. Negative samples (non-functional sites) were carefully selected and balanced with positive samples. The dataset was split into training and testing sets.
  • Dataset B: Used for training and evaluating the regulatory effect classification model (Task 2). This dataset also comes from the PTMint database and includes experimentally validated PPI records with specified regulatory effects (enhancement or inhibition) linked to phosphorylation sites.

Additionally, the authors used the Betts dataset as an independent benchmark for evaluating regulatory effect classification and a custom Dset_308 dataset for further generalization testing.

Key Insights and Ideas

  • PLMs are Game-Changers: Protein Language Models (PLMs) like ProtBERT and ESM-2 significantly outperform traditional biological features (like physical properties and sequence encoding) in capturing essential information for predicting phosphorylation sites. This highlights the power of self-supervised learning in understanding protein sequences.
  • Ensemble Power: Combining a CNN (for local features) and a Transformer (for global dependencies) in an ensemble model (DeepPhosPPI-2) is crucial for accurately predicting the regulatory effects of phosphorylation on PPIs. Neither architecture alone is sufficient.
  • Soft Voting Wins: Among different ensemble strategies (soft voting, hard voting, max confidence voting), soft voting proved to be the most effective for integrating the predictions of the CNN and Transformer models, leading to robust performance, especially in imbalanced datasets.
  • Context is King: The "STY-Sliding Window" approach, which focuses on residues surrounding S, T, and Y sites, is vital for capturing the local context that influences phosphorylation's regulatory impact on PPIs.

Key Results

  • Task 1 Performance: DeepPhosPPI-1 (using ESM-2 embeddings) achieved superior AUROC (0.807) and AUPR (0.818) scores compared to the state-of-the-art method PhosPPI-1. It also outperformed PhosPPI-1 in Accuracy (0.745 vs. 0.719) and Recall (0.755 vs. 0.690), demonstrating better overall predictive performance and a stronger ability to identify functional sites.
  • Task 2 Performance: DeepPhosPPI-2 (using ProtBERT embeddings) achieved the best overall performance on the Betts benchmark dataset, with Accuracy of 0.775, AUROC of 0.820, and AUPR of 0.921, significantly outperforming competing methods like PhosPPI-2, Betts, and FoldX. Similar strong performance was observed on the independent Dset_308 test set.
  • Ablation Studies: The ensemble model DeepPhosPPI-2 consistently outperformed its individual components (AttCNN-PhosPPI and Transformer-PhosPPI), confirming the benefit of integrating both local and global feature extraction strategies.

Limitations and Future Work

  • PLM Choice: While DeepPhosPPI-2 (ProtBERT) generally outperformed DeepPhosPPI-2 (ESM-2), the authors note that the performance difference might stem from variations in how the PLMs represent sequences and their compatibility with the ensemble architecture. This suggests that further exploration of different PLMs could yield even better results.
  • Dataset Limitations: The study relies on existing databases for training and validation. While new datasets were created, the scope and diversity of these datasets could be expanded.
  • Future Directions: The authors highlight the framework's flexibility in substituting feature embedding modules, suggesting its adaptability for various downstream applications and emerging techniques in biological sequence analysis. The potential for broader applications in predicting other PTM effects on PPIs or other biological processes is also implied.

Summary updated

BibTeX
@article{gong2025deepphosppi,
  author = {Yinyin Gong and Rui Li and Yan Liu and Jilong Wang and Danny Z. Chen and Chee Keong Kwoh},
  title = {{DeepPhosPPI: a deep learning framework with attention-CNN and transformer for predicting phosphorylation effects on protein-protein interactions}},
  journal = {Briefings in Bioinformatics},
  year = {2025},
  doi = {10.1093/bib/bbaf462},
}
Figure