Multi-modal features-based human-herpesvirus protein-protein interaction prediction by using LightGBM

Xiaodi Yang, Stefan Wuchty, Zeyin Liang, Li Ji, Bingjie Wang, Jialin Zhu, Ziding Zhang, Yujun DongBriefings in Bioinformatics 2024 · DOI

TL;DR

This paper introduces a novel multi-modal approach using LightGBM to predict protein-protein interactions (PPIs) between human and herpesviruses. By integrating sequence, network, and functional data through advanced embedding techniques, their method significantly outperforms existing approaches, offering a more robust and accurate tool for understanding viral infection mechanisms.

What Problem Does This Paper Address?

Understanding how human and herpesvirus proteins interact (PPIs) is crucial for deciphering viral infection mechanisms, especially in immunocompromised individuals where herpesvirus reactivation can lead to severe diseases. While experimental methods exist, they are time-consuming and labor-intensive. Computational prediction methods are essential, but existing approaches for human-herpesvirus PPIs are limited. Many current methods rely on single data modalities (like protein sequences) or traditional feature encoding techniques, which may not fully capture the complex nature of these interactions. Furthermore, the evaluation of these models often uses datasets that aren't rigorously partitioned, potentially leading to overestimated performance and an inability to predict interactions with novel viral proteins or across different herpesvirus subtypes. This paper aims to address these limitations by developing a more comprehensive and rigorously evaluated prediction model.

Methods

The core of this work lies in a multi-modal feature fusion strategy combined with the LightGBM machine learning algorithm.

  • Multi-modal Feature Extraction:
    • Sequence: Protein sequences are encoded using Doc2vec, a document embedding technique that treats protein sequences as "sentences" and amino acid k-mers as "words." This captures semantic features and contextual information within sequences. A novel k-mer extraction method (k=1) was found to significantly improve performance.
  • Network: Protein interaction networks are represented using Net2vec, a graph embedding technique. This involves generating node sequences via random walks on protein interaction networks and then feeding these sequences into a Word2vec model to obtain feature vectors for each protein.
  • Function: Gene Ontology (GO) annotations are used to represent functional properties. GO2vec employs a similar Net2vec approach on a GO hierarchical network and GO-protein annotation data to generate functional feature vectors.
  • Model Training: The extracted multi-modal features (sequence, network, and function) are concatenated to form a comprehensive feature vector for each protein pair. This fused feature set is then used to train a LightGBM classifier. LightGBM, an efficient gradient boosting decision tree algorithm, was chosen for its speed and high accuracy.
  • Rigorous Evaluation: The authors emphasize the importance of rigorous dataset partitioning. They constructed both non-rigorous and rigorous benchmarking datasets. The rigorous partition ensures that viral proteins in the test set are sequence-dissimilar to those in the training set, providing a more realistic assessment of the model's ability to predict interactions with novel viral proteins.

Datasets

The study utilized several public molecular interaction databases for data collection:

  • Training and Evaluation Data: IntAct, BioGRID, HPIDB, VirHostNet, and VirusMentha were used to compile human-herpesvirus PPI data.
  • Negative Sample Generation: A "Dissimilarity-Based Negative sampling" method was employed, and negative samples were generated to be 10 times the number of positive samples.
  • Benchmarking Datasets:
    • Non-rigorous partition: Random sampling of training and test sets.
  • Rigorous partition: Specifically designed to ensure sequence dissimilarity between viral proteins in the training and test sets.
  • External Datasets for Comparison: The authors also used datasets from previous studies for comparison, including those used by LSTM-PHV and DeepViral.

Key Insights and Ideas

  • Multi-modal Fusion is Key: Combining sequence, network, and functional information provides a significantly more powerful prediction capability than any single modality alone. This suggests that a holistic view of protein characteristics is essential for accurate PPI prediction.
  • Novel k-mer Extraction Boosts Performance: The discovery that using single amino acid residues (k=1) as "words" in the Doc2vec model dramatically improved sequence-based prediction performance was a significant finding, highlighting the importance of exploring different feature extraction granularities.
  • Rigorous Evaluation Matters: The use of rigorously partitioned datasets revealed performance differences that might be masked by less stringent evaluation methods, underscoring the need for robust validation to assess true generalization ability.
  • Transfer Learning Potential: The model's ability to predict human-cytomegalovirus PPIs even when trained on a dataset without cytomegalovirus data suggests its capacity to generalize across different herpesvirus subtypes.

Key Results

  • The multi-modal LightGBM model consistently outperformed single-modal models (sequence, network, function) across both non-rigorous and rigorous datasets, with AUPRC scores being 3.5 and 2.2 percentage points higher, respectively.
  • The proposed method significantly outperformed traditional feature encoding-based ML methods (DPC, CT, AC with RF/SVM/MLP) and existing state-of-the-art human-virus PPI prediction methods like LSTM-PHV and DeepViral, demonstrating superior AUROC and AUPRC values.
  • In a cross-viral subtype prediction test, the multi-modal model showed better performance compared to DeepViral, indicating its enhanced ability to predict interactions across different herpesvirus subtypes.
  • The model successfully predicted several experimentally confirmed human-herpesvirus interactions, including those involving EBV and CMV proteins, with high confidence under a low false positive rate.
  • Functional and pathway enrichment analysis of predicted targets revealed common, related, and unique enriched terms across different herpesvirus subtypes, providing insights into their specific infection mechanisms and host interactions.

Limitations and Future Work

  • Generalization to Unseen Proteins: The authors acknowledge that existing methods, including theirs, can still suffer from limitations in generalizing to interactions involving entirely new proteins not seen during training.
  • Negative Sample Selection: The challenge of selecting appropriate negative samples remains a critical issue that can impact prediction accuracy and generalization.
  • Integration of Structural Data: The paper suggests that incorporating protein structural information, potentially derived from tools like AlphaFold2, could further enhance prediction models.
  • Leveraging Large Language Models: The authors propose that large language models (LLMs) trained on vast protein sequence data could be powerful for learning protein sequence patterns and improving human-virus PPI prediction.

Summary updated

BibTeX
@article{yang2024multi,
  author = {Xiaodi Yang and Stefan Wuchty and Zeyin Liang and Li Ji and Bingjie Wang and Jialin Zhu and Ziding Zhang and Yujun Dong},
  title = {{Multi-modal features-based human-herpesvirus protein-protein interaction prediction by using LightGBM}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  doi = {10.1093/bib/bbae005},
}
Figure