Network-based virus-host interaction prediction with application to SARS-CoV-2

Hangyu Du, Feng Chen, Hongfu Liu, Pengyu Hong

TL;DR

This paper introduces a novel machine learning framework, IMSP, that builds a multi-layer network of virus-host interactions to predict potential infection targets and protein-level interactions for new viruses like SARS-CoV-2. By leveraging existing knowledge of related coronaviruses, IMSP successfully identified five potential host animals for SARS-CoV-2 and 19 key interactions within the human innate immune pathway, offering valuable insights for pandemic preparedness and therapeutic development.

What Problem Does This Paper Address?

The COVID-19 pandemic, caused by SARS-CoV-2, highlighted a critical challenge: understanding how a novel virus interacts with its hosts at both the organism and protein levels. While significant research has focused on SARS-CoV-2, much of it has been protein-specific or provided only high-level overviews. Furthermore, a vast amount of data and knowledge about other highly relevant RNA viruses and their host interactions remains underutilized. This gap hinders our ability to predict potential hosts, understand viral invasion mechanisms, and develop effective treatments and vaccines for emerging infectious diseases. The authors aim to bridge this gap by developing a method that can predict these complex interactions by learning from existing, related viral data.

Methods

The core of this work is the Infection Mechanism and Spectrum Prediction (IMSP) framework. IMSP operates in two main stages:

  1. Virus-Host Interaction Network Construction: A multi-layer, heterogeneous network is built. This network includes an "organism layer" (viruses and hosts) and a "protein layer" (virus proteins and host proteins). Edges represent four types of relationships: virus-host infection, protein-protein interaction (PPI), protein-homolog similarity, and organism-protein belonging. This network integrates existing knowledge from various sources, including experimental data and literature.
  2. Network-Based Representation Learning and Prediction: IMSP employs a machine learning approach to learn representations of these interactions.
  3. Structural Embeddings: A graph representation learning technique (specifically, a modified Node2vec) is used to capture the network's structural features. This involves learning embeddings for nodes based on their connectivity and edge weights.
  4. Content Embeddings: Textual information about nodes (e.g., protein names, functions) is processed using a Word2vec-based model (Text2vec) to generate content embeddings.
  5. Full Edge Embeddings: Structural and content embeddings are combined to create a comprehensive representation for each potential edge (interaction).
  6. MLP Classifier: A Multi-layer Perceptron (MLP) classifier is trained on these edge representations to predict various interaction types, including infection, PPI, and non-interactions.

This approach allows IMSP to predict novel interactions at both the protein (mechanism) and organism (spectrum) levels.

Datasets

The study utilized existing datasets and knowledge bases for training and evaluation. These included:

  • Virus-Host Interaction Data: Information on known virus-host infections and protein-protein interactions (PPIs) for seven human coronaviruses (including SARS-CoV-2) and 17 host organisms.
  • Protein Homolog Similarity Data: Data from NCBI BLASTP results to determine pairwise similarity between protein homologs.
  • Protein Function Data: Information on the molecular functions of proteins.

The paper does not introduce a new benchmark dataset but rather constructs a comprehensive network from existing sources.

Key Insights and Ideas

  • Leveraging Cross-Viral Knowledge: The central idea is that knowledge gained from well-studied viruses can be effectively transferred to predict interactions for novel viruses like SARS-CoV-2. This is crucial when experimental data for new threats is scarce.
  • Multi-Layered Network Approach: Building a network that integrates both organism-level and protein-level information provides a richer context for prediction than single-layer approaches.
  • Predicting Beyond Known Interactions: IMSP goes beyond simply cataloging known interactions; it actively predicts novel and highly possible interactions, offering new hypotheses for experimental validation.
  • Innate Immune Pathway Significance: The focus on the innate immune pathway is critical, as viruses often target these systems to evade detection and replicate. Identifying these interactions can reveal key vulnerabilities.

Key Results

  • Potential Infection Targets: IMSP predicted five mammalian hosts that are highly likely to be susceptible to SARS-CoV-2 infection. These include rats, sheep, camels, swine, and squirrels.
  • Innate Immune Pathway Interactions: The model identified 19 highly probable interactions between SARS-CoV-2 proteins and human proteins involved in the innate immune pathway. These interactions are hypothesized to play roles in suppressing the host's immune response.
  • Performance Metrics: IMSP demonstrated superior performance compared to five baseline models in link prediction tasks. It achieved an overall accuracy of 97.1% and a weighted F1-score of 0.971, significantly outperforming other methods. For infection predictions, it achieved an F1-score of 0.854, a 40.4% increase over the second-best model.

Limitations and Future Work

  • Experimental Validation: The authors acknowledge that their predictions require validation through wet-lab experiments, which are resource-intensive and were not feasible within the scope of this study.
  • Data Quality and Bias: The accuracy of predictions is dependent on the quality and completeness of the input data. Potential biases (sampling, representation, population) exist due to the selective inclusion of proteins and the availability of data for certain organisms.
  • Future Directions:
    • Incorporating gene set enrichment and sequence motif analysis to further improve model performance.
  • Expanding the network with more data as it becomes available to mitigate biases.
  • Applying the IMSP pipeline to investigate other emerging viruses and their host interactions.

Summary updated

Abstract

COVID-19, caused by Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2), has quickly become a global health crisis since the first report of infection in December of 2019. However, the infection spectrum of SARS-CoV-2 and its comprehensive protein-level interactions with hosts remain unclear. There is a massive amount of underutilized data and knowledge about RNA viruses highly relevant to SARS-CoV-2 and proteins of their hosts. More in-depth and more comprehensive analyses of that knowledge and data can shed new light on the molecular mechanisms underlying the COVID-19 pandemic and reveal potential risks. In this work, we constructed a multi-layer virus-host interaction network to incorporate these data and knowledge. We developed a machine-learning-based method to predict virus-host interactions at both protein and organism levels. Our approach revealed five potential infection targets of SARS-CoV-2 and 19 highly possible interactions between SARS-CoV-2 proteins and human proteins in the innate immune pathway.
BibTeX
@inproceedings{du2021network,
  author = {Hangyu Du and Feng Chen and Hongfu Liu and Pengyu Hong},
  title = {{Network-based virus-host interaction prediction with application to SARS-CoV-2}},
  booktitle = {Patterns},
  year = {2021},
  url = {https://openreview.net/forum?id=zO4H6pAyLM},
}
Figure