Uncovering the complexities of biological structures with network-based learning: An application in SARS-CoV-2

Faraz Faghri, Mike A. NallsPatterns 2021 · DOI

TL;DR

This paper highlights how network-based learning can uncover previously unknown interactions between viruses and their hosts, using SARS-CoV-2 as a case study. By constructing a comprehensive biological network, researchers can predict potential drug targets and understand viral infection mechanisms more deeply.

What Problem Does This Paper Address?

Biological systems are incredibly complex, and understanding the intricate web of interactions between different components, especially between viruses and their hosts, is a significant challenge. Traditional data structures often struggle to capture this relational information effectively. The authors address the need for methods that can leverage existing biological knowledge to predict novel interactions, which is crucial for developing treatments and vaccines, particularly in the context of rapidly evolving threats like SARS-CoV-2.

Methods

The core methodology employed is network-based learning. This approach constructs a biological network that integrates data from multiple mammals and SARS-CoV-2. The network is organized into two layers: an organism layer and a protein layer, with four types of defined interactions: protein-protein, infection, similarity (between homologs), and belonging (organism to protein).

The problem of predicting undiscovered interactions is framed as a classification task. A Multi-layer Perceptron (MLP) machine learning model is used to classify potential protein interactions into three categories: protein-protein interaction, infection, or no-interaction. This machine-learning approach differs from traditional methods by considering the comprehensive structure of the network and individual interactions, rather than relying solely on sparse, subject-level data.

Datasets

The paper describes the construction of a network using aggregated data from multiple mammals and SARS-CoV-2. The specific datasets used for training and evaluation are not explicitly detailed beyond this general description. The final virus-host interaction network comprised 246 nodes and 1,766 interactions.

Key Insights and Ideas

  • Network-based learning as a predictive tool: The fundamental idea is that by mapping known biological interactions, we can infer unknown ones, acting as a powerful hypothesis generation engine.
  • SARS-CoV-2 as a prime example: The application to SARS-CoV-2 demonstrates the immediate relevance of these methods for understanding and combating infectious diseases.
  • Beyond pairwise interactions: The network approach allows for a more holistic view, capturing complex relationships that might be missed by studying individual interactions in isolation.
  • Hypothesis generation for drug discovery: The predicted interactions can directly inform experimental validation, guiding researchers toward potential therapeutic targets.

Key Results

The application of network-based learning to SARS-CoV-2 revealed:

  • Five potential infection targets for SARS-CoV-2.
  • Nineteen possible interactions between SARS-CoV-2 proteins and human proteins.

These findings are presented as hypotheses that warrant further experimental validation.

Limitations and Future Work

The authors acknowledge several limitations and suggest future directions:

  • Data Bias: Biological networks are inherently biased by the quality and availability of existing research and funding. The inclusion of only 17 mammalian species due to data availability highlights this.
  • Data Quality and Certainty: Results can be skewed by the quality of input data, the level of certainty in existing knowledge, and missing data percentages.
  • Interpretability: Making the models more understandable for biologists is an ongoing challenge.
  • User Interaction: Developing effective ways for users to interact with the network and feed their insights back into the model is a key research direction.
  • Demographic Diversity: Incorporating demographic diversity (age, gender, racial/ethnic composition) into models is important.
  • Predicting Rare Interactions: Developing methods to predict interactions for which there is very little training data remains a challenge.
  • Robustness and Generalizability: Further work is needed to evaluate the robustness and generalizability of these models across different biological settings and domains.
  • Reproducibility: The authors emphasize the importance of transparency in network building and analysis, making their code available to ensure reproducibility and facilitate extension of their work.
  • Expanding Data Modalities: Future improvements could involve incorporating more data modalities, such as gene set enrichment analysis and sequence motif analysis.
  • Broader Applications: The approach can be extended to domains beyond infectious diseases by appropriately redefining network nodes and interactions.

Summary updated

BibTeX
@article{faghri2021uncovering,
  author = {Faraz Faghri and Mike A. Nalls},
  title = {{Uncovering the complexities of biological structures with network-based learning: An application in SARS-CoV-2}},
  journal = {Patterns},
  year = {2021},
  doi = {10.1016/j.patter.2021.100259},
}
Figure