Deciphering the code of viral-host adaptation through maximum entropy models

Andrea Di Gioacchino, Benjamin D. Greenbaum, Rémi Monasson, Simona CoccobioRxiv 2023 · DOI

TL;DR

This paper introduces a novel Maximum Entropy Nucleotide Bias (MENB) model that accurately predicts a virus's host and family based on its genomic nucleotide usage. This interpretable model outperforms deep learning approaches and can even be used to synthetically evolve viral genomes, offering new avenues for vaccine design and pathogen surveillance.

What Problem Does This Paper Address?

Understanding how viruses adapt to new hosts is crucial, especially with the rise of emerging infectious diseases like SARS-CoV-2. This adaptation involves changes in viral genomes, particularly in nucleotide usage patterns, which can be influenced by host immune systems, cellular machinery, and evolutionary pressures. The challenge lies in deciphering these complex genomic "codes" to identify a virus's origin and track its evolutionary trajectory. Existing methods often rely on sequence alignment or complex deep learning models, which can be computationally expensive and lack interpretability. This paper aims to bridge this gap by developing a more interpretable and efficient model for inferring host-pathogen relationships from genomic data.

Methods

The core of this work is the Maximum Entropy Nucleotide Bias (MENB) model. This is a statistical physics-inspired approach that learns the probability distribution of nucleotide usage (single, di-, and tri-nucleotides) within viral genomes. The model is trained to reproduce observed k-mer frequencies from a given host's viral sequences.

Key aspects of the method include:

  • Maximum Entropy Principle: This principle guides the model to find the simplest distribution that explains the observed data, avoiding overfitting and promoting interpretability.
  • Learnable "Forces": The model infers parameters, termed "forces," which represent the selective pressures acting on specific nucleotide motifs. These forces are directly interpretable as biological influences.
  • Generative Capabilities: The MENB model can generate synthetic viral sequences by manipulating these forces, allowing researchers to explore hypothetical evolutionary scenarios (e.g., reducing CpG content without altering protein sequence).
  • Comparison with Deep Learning: The authors compare MENB's performance against VIDHOP, a state-of-the-art deep neural network, demonstrating MENB's competitive accuracy with significantly reduced computational cost and improved interpretability.

Datasets

The study utilized viral sequences from the BV-BRC database. For training and evaluation, they focused on sequences from three host classes: human, avian, and swine. Specifically, they selected at least 150 viral genomes for each host class, covering four viral families: Coronaviridae, Flaviviridae, Picornaviridae, and Orthomyxoviridae (specifically Influenza A).

For the SARS-CoV-2 analysis, they used data from Nextstrain, a real-time pathogen surveillance platform.

Key Insights and Ideas

  • Host and Viral Family Leave Distinct Signatures: The MENB model successfully decodes distinct nucleotide usage patterns that are characteristic of both the viral family and the host it infects. This suggests a deep interplay between viral evolution and host environment.
  • MENB Outperforms Deep Learning in Specific Scenarios: While deep learning models are powerful, MENB demonstrates superior performance in host classification when phylogenetic information is limited, highlighting the value of its interpretable, feature-based approach.
  • Interpretable Parameters Reveal Biological Pressures: The "forces" learned by the MENB model directly correspond to biological selective pressures, such as the well-known avoidance of CpG motifs in human viruses. This interpretability is a significant advantage over black-box models.
  • Generative Power for Synthetic Biology: The ability to synthetically alter viral genomes while preserving protein sequences opens doors for designing attenuated viruses for vaccines or understanding the functional impact of specific nucleotide biases.
  • Host Adaptation as a "Molecular Clock": The log-probability difference metric used to track adaptation over time in H1N1 influenza acts like a molecular clock, quantifying the virus's evolutionary journey in a new host.

Key Results

  • Host and Viral Family Classification: MENB models achieved high accuracy in classifying both the host and viral family of unknown viral sequences, outperforming state-of-the-art deep neural networks in certain settings. For instance, the "MENB-H,V" strategy achieved ~97% accuracy in predicting both host and viral family.
  • Computational Efficiency: MENB models are significantly faster to train and run compared to deep learning models like VIDHOP. MENB requires minutes on a few cores, while VIDHOP can take hours on many cores.
  • Generative Model Validation: The model successfully generated synthetic sequences with altered nucleotide composition (e.g., reduced CpG content) without changing the encoded protein sequence, demonstrating its utility for synthetic biology applications.
  • Tracking Host Adaptation: Analysis of H1N1 influenza and SARS-CoV-2 showed that MENB can effectively track viral adaptation to new hosts over time, revealing the dynamics of selective pressures. For SARS-CoV-2, a steady adaptation to human nucleotide usage was observed, with an accelerated trend in recent variants.

Limitations and Future Work

  • Data Bias: The study acknowledges limitations due to biases in sequence data, particularly geographical and temporal biases in SARS-CoV-2 data.
  • Unknown Origin of SARS-CoV-2: For SARS-CoV-2, inferring the original host was challenging due to the lack of comprehensive data from potential reservoir species (e.g., bats, pangolins).
  • Short Time Scale for SARS-CoV-2 Adaptation: The relatively short sampling period for SARS-CoV-2 adaptation makes it harder to distinguish signals from non-synonymous mutations versus nucleotide usage changes.
  • Future Work: The authors suggest integrating MENB into metagenomic studies for host identification, using it for emerging pathogen surveillance, predicting mutations affecting immunogenicity, and assisting in RNA vaccine design by fine-tuning motif usage. They also highlight the potential for combining MENB with models of epistatic fitness landscapes to capture more complex evolutionary dynamics.

Summary updated

Abstract

Abstract Understanding how the genome of a virus evolves depending on the host it infects is an important question that challenges our knowledge about several mechanisms of host-pathogen interactions, including mutational signatures, innate immunity, and codon optimization. A key facet of this general topic is the study of viral genome evolution after a host-jumping event, a topic which has experienced a surge in interest due to the fight against emerging pathogens such as SARS-CoV-2. In this work, we tackle this question by introducing a new method to learn Maximum Entropy Nucleotide Bias models (MENB) reflecting single, di- and tri-nucleotide usage, which can be trained from viral sequences that infect a given host. We show that both the viral family and the host leave a fingerprint in nucleotide usages which MENB models decode. When the task is to classify both the host and the viral family for a sequence of unknown viral origin MENB models outperform state of the art methods based on deep neural networks. We further demonstrate the generative properties of the proposed framework, presenting an example where we change the nucleotide composition of the 1918 H1N1 Influenza A sequence without changing its protein sequence, while manipulating the nucleotide usage, by diminishing its CpG content. Finally we consider two well-known cases of zoonotic jumps, for the H1N1 Influenza A and for the SARS-CoV-2 viruses, and show that our method can be used to track the adaptation to the new host and to shed light on the more relevant selective pressures which have acted on motif usage during this process. Our work has wide-ranging applications, including integration into metagenomic studies to identify hosts for diverse viruses, surveillance of emerging pathogens, prediction of synonymous mutations that effect immunogenicity during viral evolution in a new host, and the estimation of putative evolutionary ages for viral sequences in similar scenarios. Additionally, the computational frame-work introduced here can be used to assist vaccine design by tuning motif usage with fine-grained control. Author summary In our research, we delved into the fascinating world of viruses and their genetic changes when they jump from one host to another, a critical topic in the study of emerging pathogens. We developed a novel computational method to capture how viruses change the nucleotide usage of their genes when they infect different hosts. We found that viruses from various families have unique strategies for tuning their nucleotide usage when they infect the same host. Our model could accurately pinpoint which host a viral sequence came from, even when the sequence was vastly different from the ones we trained on. We demonstrated the power of our method by altering the nucleotide usage of an RNA sequence without affecting the protein it encodes, providing a proof-of-concept of a method that can be used to design better RNA vaccines or to fine-tune other nucleic acid-based therapies. Moreover the framework we introduce can help tracking emerging pathogens, predicting synonymous mutations in the adaptation to a new host and estimating how long viral sequences have been evolving in it. Overall, our work sheds light on the intricate interactions between viruses and their hosts.
BibTeX
@misc{gioacchino2023deciphering,
  author = {Andrea Di Gioacchino and Benjamin D. Greenbaum and Rémi Monasson and Simona Cocco},
  title = {{Deciphering the code of viral-host adaptation through maximum entropy models}},
  year = {2023},
  doi = {10.1101/2023.10.28.564530},
  howpublished = {bioRxiv},
}
Figure