In February 2021, it was reported that 7 Russian poultry farm workers were infected with H5N8 avian influenza. This avian influenza subtype has never infected humans before, and the virus's gene sequence is quickly uploaded to the gene data repository GISAID.
This provides an opportunity for Georgetown University biologist Colin Carlson of Washington, DC. “I immediately thought, ‘I want to run it through FluLeap’,” he said.
FluLeap is a machine learning algorithm that uses sequence data to classify influenza viruses as avian influenza virus or human virus. The model has been trained on a large number of influenza genomes to understand the differences between genomes that infect humans and infected birds.
But the model has never seen the H5N8 virus classified as human, and Carlson is curious about its effect on this new subtype.
a little surprisingly, the model identifies it as human with a confidence of 99.7%. Rather than simply reaffirming patterns in its training data, such as the fact that the H5N8 virus usually does not infect humans, the model seems to infer some biology features of human compatibility.
"Surprisingly, this model worked," Carlson said. "But it's a data point; if I could do it a thousand more times, it would be even more amazing. The zoonotic process of viruses spreading from wildlife to humans has led to most epidemics. With climate change and human encroachment of animal habitats increasing the frequency of these events, understanding zoonotic diseases is crucial to efforts to prevent epidemics, or at least to be better prepared.
Researchers estimate that about 1% of mammalian viruses on Earth have been identified. So some scientists are trying to expand our understanding of this global viral group by sampling wildlife.
It's a difficult task But over the past decade or so, a new discipline has emerged—researchers use statistical models and machine learning to predict various aspects of the disease’s emergence, such as global hotspots, possible animal hosts, or the ability of a specific virus to infect humans. Advocates of this “zoonotic risk prediction” technique, which believes it will allow us to better target monitoring targets to the right areas and conditions, and guide the development of vaccines and treatments that are most likely to be needed.
However, some researchers are skeptical of the ability of predictive technologies to cope with the size and changing nature of the viral community. Efforts to improve models and the data they rely on are underway, but if this Some tools to mitigate future pandemics, they will need to be part of a broader effort.
Virus Search
Some researchers have long believed that expanding our understanding of virus diversity will help manage pandemic threats. PREDICT is a $200 million project funded by United States Agency for International Development ( USAID) that spent about a decade looking for animal viruses. By the end of 2020, it found 949 new viruses in samples of wildlife, livestock and humans in 34 countries.
In hindsight, some of the findings from PREDICT seem prescient. 2017 A year-old study estimated that there were thousands of undetected coronaviruses in bats and predicted that Southeast Asia would be the home of the most virus-numbered family of SARS-CoV-2. It also linked activities involving high levels of human-wildlife contact with coronavirus prevalence.
1
10
17 Another study in 2017 gathered data on which viruses infect which mammals, creating a database of virus-host associations. “The goal is to understand which viruses can infect humans, which animals we most often infect new viruses and the underlying factors driving these patterns,” Kevin, an ecologist and head of research at the New York City Ecological Health Alliance. Olival said it is a nonprofit focused on biomonitoring and conservation.Analysis by the
research team showed that the proportion of viruses that can infect humans in a particular host species is affected by the close relationship between humans and factors that affect human contact with wildlife, such as population density and the degree of urbanization within the geographical scope of the species.
The team used statistical models to predict groups and regions of animal populations and regions that may carry large numbers of undetected viruses - bats, as well as rodents and primates, have prominent positions in areas such as South America, Africa and Southeast Asia. The researchers also found that the characteristics associated with the virus are zoonotic, such as the range of species it can infect.
The team said that this information can help guide monitoring efforts. “It enables us to predict the most risky areas,” said Jonna Mazet, an epidemiologist at the University of California, Davis, who directed PREDICT.
Identification of specific threats also enables local researchers and health care workers to adjust their mitigation and response capabilities. “It allows the community to say 'we have this, this and this, and we can reduce the risk in this way',” Mazet said.
PREDICT is just a pilot project. “It generates a lot of data, but it’s just a drop in the bucket,” Olival said. "We need something bigger." Therefore, in 2016, the Global Virus Project (GVP), which is considered a government agency, NGO and the researchers' Global Partnership , aiming to discover most viruses in mammals and birds (most zoonotic virus origin).
However, in the face of criticism from some researchers, it has never received funding. Mazet said it exists today as a nonprofit organization to provide countries with the knowledge they need to conduct their own virus investigations. USAID launched a smaller, much lower-cost project in October 2021 called “Discovering and Exploring Emerging Pathogens – Viral Zoonosis (DEEP VZ). One criticism of GVP for GVP is that the size of the task is simply unmanageable. Forecasting researchers estimate 1.67 million unknown viruses in 4 mammals and birds, and despite this number controversial, there is no doubt that the virions are huge.
It is also changing, so one-time discovery work is not enough. “RNA virus evolves at a very high rate,” said Edward Holmes, a virologist at the University of Sydney, Australia. “So you have to keep doing that.” "
There are also doubts about whether the project will spot a potential epidemic." I have no problem with it in understanding virus evolution and ecology ," said Sherlock . "But as a prediction tool to understand what will happen next, it is impossible.
One problem is that some host species and virus families have been studied in depth, but others have been hardly touched. Existing data also tends to overflow viruses. So far, most of the predictions have been based on “completely biased data,” says Jemma Geoghegan, a virologist at the University of Otago in New Zealand.
Furthermore, even if the virus is discovered and its genome is sequenced, many factors that may affect its potential to trigger a pandemic, such as its ability to infect humans and spread between people, are still unclear. “And then you have to do all these experiments, and it will take years and cost a lot of money,” Sherlock Holmes said.
This is where machine learning may provide shortcuts. Rather than trying to fully characterize each new virus, use a model to mark high priority targets for further investigation. "What we need is a downstream classification system so we know which viruses need to be characterized by in-depth virology studies," said Sara Sawyer, a virologist at the University of Colorado Boulder, .
model internal
When a virus is discovered, it is usually very little known about it except its gene sequence. Therefore, models that use only viral genomes to classify viruses will be particularly useful.Computational virologist Nardus Mollentze at the University of Glasgow, UK, and colleagues have developed a model like this that evaluates viruses in part by using measurements of genetic similarity to the human genome. Evolutionary stress of
viruses may lead to similarity to gene fragments in the host genome - escaping the innate immune system or helping replication. When tested on a viral library containing 861 known viruses, the algorithm can classify them as zoonotic or nonzoonotic viruses with a 70% accuracy.
After that, Mollentze joined the Virus Adventure Research Project (Verena), a consortium of researchers seeking to develop and improve zoonotic prediction models.
Mollentze works with Verena researchers to combine his algorithm with techniques that exploit knowledge of which viruses infect which hosts, including methods to infer unknown host-viral associations. This combination approach improves performance by approximately 10 percentage points 7. In the future, knowledge about how viruses interact with hosts at the molecular level can be incorporated.
"It will be about protein and biochemistry ," said Carlson, who directed Verena. "This is the future of this."
An important goal is to understand which models work well and why. Some models are classified only based on patterns in the data, and some can infer the reasons for these patterns, but it is difficult to distinguish them. "There is a question: are we just teaching machines to reiterate what they already know, or are learning the principles of bringing into a new space?" Carlson said.
To make progress, the process of validating the model is crucial. For example, some studies have tried to predict which species hosts zoonotic viruses, with mixed results, but few systematic comparisons, so it is difficult to know which approach works.
To solve this problem, in early 2020, Verena researchers used predicting which bat species may carry beta coronavirus as a case study. They created eight statistical models and used them to generate a list of suspicious hosts. In the next 16 months, 47 new bat hosts were found.
When the researchers compared these with their predictions, they found that half of the models performed significantly better than accidentality. These models include characteristics such as the lifespan or size of the species. The other four models did not take these features into account and performed poorly.
data development
Any artificial intelligence (AI) algorithm is fundamentally limited by its input data. “Artificial intelligence works when an algorithm is trained on a large amount of quality data,” Sawyer said. "But only a small amount of spillovers occur every year, and virus data tend to be dirty and there is a lot of information missing. Most researchers think the data is insufficient at the moment. "We don't have enough high-quality data to do the predictions well," Mazet said.
To some extent, modeling relies on scientists to collect new data, but so far, virus discovery work has been based on considerations such as the highest risk locations and situations. What modelers really need is sampling aimed at improving geographic and classification coverage.
provides models with more of this kind of data that changes the horizon of questions that can be asked. "With a million data points, you can show how deforestation increases virus prevalence in bats," Carlson said. "With a trillion points, you can predict spillovers like the weather."
To get closer to this, global cooperation is required, with open data sharing as the norm, and everyone abides by data standards. The obstacles in this regard are more politics, culture and ethics than science. For example, academic incentives surrounding publications are a barrier to rapid data sharing. It is also crucial that countries that share genetic data benefit from it.
"This is the key issue, and dealing with it involves building trust," Olival said. “Make sure you not only have to use the vaccine, but also give back to society through training, capacity building and co-authoring of papers."
International Treaty, which came into force in 2014, " Nagoya Protocol " stipulates the sovereignty of countries over natural resources, including biological samples, and allows them to require a benefit-sharing agreement in exchange for obtaining such samples.
However, some laboratories can now synthesize pathogens or start developing vaccines using only gene sequencing data. "We do not have any provisions in international law that involve sequence data," Carlson said. "Nagoya was not born for that world.
" Similar questions may one day apply to zoonotic risk predictions. "We are using data collected by researchers in the global South," Carlson said. "There are reasonable questions about what it means to get this data and develop technology."
prediction and preparation
In order for modeling to have a practical impact, it must lead to publicly accessible tools that provide actionable, locally relevant information. Modeling also needs to be better combined with experimental work to ask about the characteristics of the pathogen.
Just as the model may tag candidate viruses for further research, these investigations may also generate information that can be used to validate and optimize the model. However, interdisciplinary exchanges are currently restricted. “These communities don’t talk much, or even read each other’s papers,” Sawyer said.
Modelers also need to clearly convey the uncertainty inherent in their work and what they say is the prediction so that they do not over-examination of the benefits. “No one said we will have the exact time, place and species that will lead to the next pandemic,” Olival said. Researchers are dealing with the probability, unexpected things can and do happen.
Even in the best case scenario, prediction tools cannot completely prevent the outbreak. “I absolutely don’t think we should tie the world security to these models,” Carlson said.
However, their value is obvious with the improvement of global monitoring systems, targeted vaccine development, and efforts to build health care capacities around the world. “They let us do two things: understand what’s going on around us and prioritize,” Carlson said. Ultimately, this may help reduce the frequency of epidemics. “We can better prevent some of these,” Carlson said. “But it requires us to do better in what we are doing.
1.Carlson, C. J. et al. Phil. Trans. R. Soc. London. B 376, 20200358 (2021).
2.Anthony, S. J. et al. Virus Evol. 3, vex012 (2017).
3.Olival, K. et al. Nature 546, 646–650 (2017).
4.Carroll, D. et al. Science 359, 872–874 (2018).
5. Wille, M., Geoghegan, J. L. & Holmes, E. C. PLoS Biol. 19, e3001135 (2021).
6.Mollentze, N., Babayan, S. A. & Streicker, D. G. PLoS Biol. 19, e3001390 (2021).
7.Poisot, T. et al. Preprint at https://arxiv.org/abs/2105.14973 (2022).
8.Becker, D. J. Lancet Microbe 3, E625–E637 (2022).
