Predicting hosts based on early SARS-CoV-2 samples and analyzing the 2020 pandemic

Qian Guo, Mo Li, Chunhui Wang, Jinyuan Guo, Xiaoqing Jiang, Jie Tan, Shu‐Fang Vivienne Wu, Peihong Wang, Tingting Xiao, Man Zhou, Zhencheng Fang, Yonghong Xiao, Huaiqiu ZhuScientific 2021 · DOI

Abstract

The SARS-CoV-2 pandemic has raised concerns in the identification of the hosts of the virus since the early stages of the outbreak. To address this problem, we proposed a deep learning method, DeepHoF, based on extracting viral genomic features automatically, to predict the host likelihood scores on five host types, including plant, germ, invertebrate, non-human vertebrate and human, for novel viruses. DeepHoF made up for the lack of an accurate tool, reaching a satisfactory AUC of 0.975 in the five-classification, and could make a reliable prediction for the novel viruses without close neighbors in phylogeny. Additionally, to fill the gap in the efficient inference of host species for SARS-CoV-2 using existing tools, we conducted a deep analysis on the host likelihood profile calculated by DeepHoF. Using the isolates sequenced in the earliest stage of the COVID-19 pandemic, we inferred that minks, bats, dogs and cats were potential hosts of SARS-CoV-2, while minks might be one of the most noteworthy hosts. Several genes of SARS-CoV-2 demonstrated their significance in determining the host range. Furthermore, a large-scale genome analysis, based on DeepHoF's computation for the later pandemic in 2020, disclosed the uniformity of host range among SARS-CoV-2 samples and the strong association of SARS-CoV-2 between humans and minks.

Introduction

Identifying the host range of a novel virus—the species it can infect—is a critical step in managing outbreaks. Traditional methods often rely on sequence similarity (e.g., BLAST) or phylogenetic family trees. However, these methods struggle with "dark matter" viruses that lack close relatives in existing databases. Furthermore, sequence similarity does not always guarantee functional similarity in host infection. There is a clear need for a tool that can automatically extract evolutionary signatures from raw genomic data to predict host likelihood without requiring a perfect match in the existing catalog.

Methods

We developed DeepHoF (Deep learning-based Host Finder) using a BiPath Convolutional Neural Network (BiPathCNN). The model processes viral genomes through two parallel paths: one for nucleotide sequences and one for codon sequences, allowing it to capture both raw genetic patterns and protein-coding information. Unlike previous models that required handcrafted features, DeepHoF learns relevant genomic features automatically.

The model outputs likelihood scores for five host types: plant, germ, invertebrate, non-human vertebrate, and human. To ensure statistical robustness, a p-value system determines if a host prediction is significantly distinct from non-infectious background data.

Datasets

The model was trained on 55,283 viral genomes released before 2018. Performance was validated on a chronologically separate test set (2018–2019). For SARS-CoV-2 analysis, we used the 17 earliest sequenced isolates from December 2019 and a retrospective dataset of over 53,000 genomes collected through September 2020.

Results

DeepHoF demonstrated superior performance, achieving an AUC of 0.975 across the five host classes, significantly outperforming BLAST-based methods (AUC 0.800). The model's high precision (0.967) makes it a reliable tool for identifying human-infective viruses.

Figure 1: DeepHoF performance and phylogenetic analysis. (a) ROC curves comparing DeepHoF and BLAST. (b) Comparison of human host likelihood scores. (c, d) Phylogenetic analysis of whole genomes and S genes.

SARS-CoV-2 Host Identification

Using the earliest 17 sequences, DeepHoF identified humans and non-human vertebrates—specifically minks, bats, dogs, and cats—as high-probability hosts. Notably, the host likelihood profile of SARS-CoV-2 was most similar to viruses infecting minks, suggesting their susceptibility long before large-scale outbreaks were reported in mink farms.

Analysis of individual genes showed that the Spike (S) gene, ORF1ab, and ORF7b were primary drivers for predicting human infection. Interestingly, specific functional regions were more influential than overall gene length.

Pandemic Stability and Transmission

Despite mutations throughout 2020, host likelihood profiles remained tightly clustered, indicating that the virus's fundamental host preference did not shift significantly during the first year. Additionally, analysis of Dutch samples provided computational evidence for bidirectional transmission ("spill-back") between minks and humans.

Discussion

DeepHoF provides a robust framework for predicting viral host range from genomic data. While the current model does not distinguish between reservoir hosts and incidental hosts, it offers a significant advancement over similarity-based methods. Future work will focus on incorporating host-side data, such as receptor structures, and adapting the tool for metagenomic environmental scanning.

Summary updated

Abstract

The SARS-CoV-2 pandemic has raised concerns in the identification of the hosts of the virus since the early stages of the outbreak. To address this problem, we proposed a deep learning method, DeepHoF, based on extracting viral genomic features automatically, to predict the host likelihood scores on five host types, including plant, germ, invertebrate, non-human vertebrate and human, for novel viruses. DeepHoF made up for the lack of an accurate tool, reaching a satisfactory AUC of 0.975 in the five-classification, and could make a reliable prediction for the novel viruses without close neighbors in phylogeny. Additionally, to fill the gap in the efficient inference of host species for SARS-CoV-2 using existing tools, we conducted a deep analysis on the host likelihood profile calculated by DeepHoF. Using the isolates sequenced in the earliest stage of the COVID-19 pandemic, we inferred that minks, bats, dogs and cats were potential hosts of SARS-CoV-2, while minks might be one of the most noteworthy hosts. Several genes of SARS-CoV-2 demonstrated their significance in determining the host range. Furthermore, a large-scale genome analysis, based on DeepHoF's computation for the later pandemic in 2020, disclosed the uniformity of host range among SARS-CoV-2 samples and the strong association of SARS-CoV-2 between humans and minks.
BibTeX
@inproceedings{guo2021predicting,
  author = {Qian Guo and Mo Li and Chunhui Wang and Jinyuan Guo and Xiaoqing Jiang and Jie Tan and Shu‐Fang Vivienne Wu and Peihong Wang and Tingting Xiao and Man Zhou and Zhencheng Fang and Yonghong Xiao and Huaiqiu Zhu},
  title = {{Predicting hosts based on early SARS-CoV-2 samples and analyzing the 2020 pandemic}},
  booktitle = {Scientific},
  year = {2021},
  doi = {10.1038/s41598-021-96903-6},
}
Figure