Predicting Virus Mutations Using Neural Networks

Compared to other pathogens, viruses reproduce and mutate quickly. This adaptation allows them to spread across many hosts and evade immune system defenses. However, it also allows them to undermine vaccines and acquire drug resistance, which complicates medical research. For example, influenza vaccines are designed to “match” a specific set of flu strains [1]. If a strain mutates and spreads, existing vaccines may not match this new strain. During the 2018-19 flu season, researchers found that the mutation of certain H3N2 flu viruses led to a mismatch, making the season’s flu vaccine less effective [2]. Their main recommendation was that next season’s vaccine should account for this new mutation. By preventing these mismatches preemptively instead of retrospectively, researchers can advance public health outcomes significantly. In other words, predicting virus mutations is a key part of disease prevention. 

As with many phenomena in medicine, viral mutations are complex, and researchers must apply increasingly sophisticated tools to study them. One example is machine learning (ML), a family of statistical models that learn trends from data. ML models are used in many disciplines to model associations and make predictions. Neural networks are a class of ML models inspired by biological neurons and synapses. Unlike other ML models, neural networks are primarily used to uncover complex patterns from large datasets. Furthermore, neural networks rely less on prior knowledge than other methods do, which can lead to the discovery of truly novel patterns. Using neural networks, scientists can work on predicting virus mutations from several angles.  

One avenue of research is to predict the location of mutations in a virus’s genetic material. In 2009, researchers used neural networks to predict when and where mutations occur in H1 flu viruses [3]. More recently, researchers used neural networks to study the Newcastle Disease Virus (NDV). Given data on NDV genetic material, their model predicted future mutations with a success rate of 75%. The authors framed their work as a “proof of concept” for future neural network-based studies of viral mutations [4]. Beyond predicting virus mutations, researchers have also predicted their effects. The Journal of Infectious Diseases published an article about using neural networks to predict which mutations in the HIV-1 virus caused resistance to lopinavir, an antiretroviral drug [5]. By combining methods to predict mutations and identify clinically significant mutations, researchers can design vaccines and drugs to anticipate future mutations.  

Another avenue is to predict quantitative metrics about mutations, as opposed to their precise behaviors. For instance, antigenic distance measures the degree of difference between strains of a virus. Mutations increase the antigenic distance between two strains, undermining the effectiveness of a vaccine. Researchers from the Krasovsky Institute of Mathematics and Mechanics in Russia developed neural networks to predict antigenic distance from a particular sequence of influenza proteins [6]. Another example is the mutation rate, the frequency of certain mutations happening. Though this information is less precise than the location of mutations, it proved useful in studying the novel SARS-CoV-2 virus, for which extensive mutation data might not become available for some time [7].  

When modeling viral mutations, neural networks have proven their flexibility in measuring the location, impact, similarity, and rate of mutations. However, neural networks are not a panacea. Computer science theoreticians have yet to derive universal principles for neural network design. As a result, ML experts often struggle to interpret the output of a neural network model [8]. The low interpretability of neural networks complicates their use in medicine, where laws and ethics demand a high standard of accountability. Yet, as medical research and ML research expand, neural networks and other models will continue to support healthcare innovations.  

References 

[1] How Well Flu Vaccines Work. January 2020. Retrieved October 12, 2020 from https://www.cdc.gov/flu/vaccines-work/vaccineeffect.htm.  

[2] Flannery B., et al. Spread of Antigenically Drifted Influenza A(H3N2) Viruses and Vaccine Effectiveness in the United States During the 2018–2019 Season. The Journal of Infectious Diseases 2020; 221: 1. DOI:10.1093/infdis/jiz543.  

[3] Yan S.-M. and Wu G. Prediction of Mutation Position, Mutated Amino Acid and Timing in Hemagglutinins from North America H1 Influenza A Virus. The Journal of Biomedical Science and Engineering 2009; 2. DOI:10.4236/jbise.2009.22021. 

[4] Salama M. A., et al. The Prediction of Virus Mutation using Neural Networks and Rough Set Techniques. EURASIP Journal on Bioinformatics and Systems Biology 2016; 10. DOI:10.1186/s13637-016-0042-0 

[5] Wang D. and Larder B. Enhanced Prediction of Lopinavir Resistance from Genotype by Use of Artificial Neural Networks. The Journal of Infectious Diseases 2003; 188: 5. DOI:10.1086/377453.  

[6] Forghani M. and Khachay M. Convolutional Neural Network Based Approach to In Silico Non-Anticipating Prediction of Antigenic Distance for Influenza Virus. Viruses 2020; 12: 1019. DOI:10.3390/v12091019.  

[7] Pathan R. K., et al. Time Series Prediction of COVID-19 by Mutation Rate Analysis using Recurrent Neural Network-Based LSTM Model. Chaos Solitons Fractals 2020; 138: 110018. DOI:10.1016/j.chaos.2020.110018.  

[8] Wolchover N. New Theory Cracks Open the Black Box of Deep Learning. September 2017. Retrieved October 13, 2020 from https://www.quantamagazine.org/new-theory-cracks-open-the-black-box-of-deep-learning-20170921/.