Sequencing meets machine learning to fight emerging pathogens: A preview

Artur Yakimovich

TL;DR

This preview highlights how machine learning (ML) is becoming an indispensable tool for analyzing the massive datasets generated by pathogen sequencing, particularly for SARS-CoV-2. A novel ML approach, COPA scores, identifies critical genomic regions associated with pathogenicity and host interaction, offering new avenues for vaccine design and variant surveillance.

What Problem Does This Paper Address?

The COVID-19 pandemic has generated an unprecedented volume of SARS-CoV-2 genomic data. While this data is invaluable for tracking pathogen evolution and understanding its spread, manually sifting through millions of sequences to identify subtle but crucial patterns is an insurmountable task. Traditional methods are too slow and lack the sensitivity to uncover complex dependencies. This paper addresses the need for advanced computational tools, specifically machine learning, to effectively analyze this deluge of genomic information and extract actionable insights for combating emerging infectious diseases.

Methods

The paper discusses a novel ML approach proposed by Park and colleagues. This method involves:

  • Metaviromic Data Collation: Combining a large dataset of human and animal coronavirus genomes, including SARS-CoV-2, MERS-CoV, and SARS-CoV.
  • Comparative Genomics: Utilizing multiple sequence alignment and knowledge of genomic regions of interest.
  • COPA (Coronavirus Pathogenicity) Scores: Developing scores for each nucleotide in the SARS-CoV-2 genome that indicate its discriminative power in identifying pathogenic coronaviruses.
  • Ensemble of ML Algorithms: Employing and comparing standard ML algorithms like random forest, support vector machines, Bernoulli naive Bayes, gradient boosting, and multi-layer perceptron to assess their ability to identify pathogenic genomes.
  • Statistical Metamodel: Integrating these ML approaches into a unified model to pinpoint "pathogenic hotspots" within the SARS-CoV-2 genome.
  • Integration with Biological Data: Combining the ML-generated hotspot map with knowledge of B and T immune-cell epitopes and known variants-of-concern.

Datasets

  • Human and Animal Coronavirus Genomes: A dataset comprising 3,665 genomes, including SARS-CoV-2, MERS-CoV, and SARS-CoV, was used for training and analysis.
  • SARS-CoV-2 Sequence Data: Millions of SARS-CoV-2 sequences have been accumulated globally, forming the basis for this type of analysis.
  • Known Variants-of-Concern: Sequences of established SARS-CoV-2 variants were used for cross-correlation analysis.

The paper does not explicitly introduce a new benchmark dataset but rather leverages existing large-scale genomic data.

Key Insights and Ideas

  • Genomic Hotspots for Pathogenicity: ML can effectively identify specific genomic regions within SARS-CoV-2 that are highly correlated with pathogenicity and interaction with host cells.
  • Beyond Simple Identification: The COPA score approach goes beyond simply identifying differences; it quantifies the discriminative power of genomic features, offering a more nuanced understanding.
  • Overlap with Host Interaction Sites: Identified pathogenic hotspots significantly overlap with the furin cleavage site and contact sites with the angiotensin-converting enzyme 2 (ACE2) receptor, underscoring the biological relevance of these ML-identified regions.
  • Differentiating Betacoronaviruses: The ML approach can identify amino acid insertions that distinguish betacoronaviruses from alpha- and gamma-coronaviruses.
  • Synergy of ML and Immunology: Combining ML-derived pathogenic regions with knowledge of immune epitopes reveals significant overlaps, particularly in the S and N proteins, suggesting these regions are critical for immune response.
  • Variant Surveillance Potential: The methodology shows promise in identifying overlaps between high-score residues and mutations found in variants like B.1.1.7 (Alpha), suggesting its utility for tracking and predicting the emergence of new variants.

Key Results

  • 2,473 Discriminative Hotspots: The ML approach identified a substantial number of discriminative hotspots across the SARS-CoV-2 genome.
  • Overlap with Key Biological Sites: These hotspots were found to overlap with the furin cleavage site and ACE2 contact sites, as well as with potential B cell epitopes in the S and N proteins.
  • Variant Correlation: Several high-score residues and mutations for the SARS-CoV-2 Alpha variant (B.1.1.7) were observed to overlap with identified COPA hotspots.

Limitations and Future Work

  • Data Coverage: The authors acknowledge that despite the immense sequencing effort, global coverage of pathogen sequence data is still limited, with a few nations and laboratories performing the lion's share. This can lead to "shortsighted" datasets.
  • Future Directions: The work suggests that ML methodologies can be instrumental in guiding the design of future SARS-CoV-2 vaccines and improving epidemiological variant surveillance. The authors envision a future where ML drives a new era of infection biology, shifting approaches from reactive to proactive.

Summary updated

BibTeX
@inproceedings{yakimovich2022sequencing,
  author = {Artur Yakimovich},
  title = {{Sequencing meets machine learning to fight emerging pathogens: A preview}},
  booktitle = {Patterns},
  year = {2022},
  url = {https://openreview.net/forum?id=O4A71ImQAz},
}
Figure