Structured illumination microscopy combined with machine learning enables the high throughput analysis and classification of virus structure

Viruses are like the Trojan horses of the biological world; they sneak their genetic code into a living cell and then hijack it, forcing that cell to produce their own viral proteins. Yet, if scientists replace the harmful genes in a virus with other genes, the virus can be transformed into a powerful tool for biology and medical science. For example, viruses can be turned into vaccines that prime the immune system to ward off future infections. Viruses could also be made to deliver the genetic code needed to repair faulty cells, and thus treat the cause of an illness from inside the body. Nevertheless, it is complicated to produce viruses like these on a large scale. The individual viruses in one batch can be very different shapes and sizes; they can also end up displaying different proteins on their outer surface – which is the part of the virus that our immune system will see first. To optimise the production of standardised viruses, scientists need a way to test the viruses throughout the manufacture process. At the moment, the best way to do this would be with electron microscopes. Yet these microscopes cannot tell exactly which proteins are in the outer surface of the virus. Also, these methods often need purified samples of virus, so cannot be used to look at the viruses until the final stage of production. Laine et al. now report a method that can test virus production at every step of the process. This new method uses a different type of microscopy called super-resolution imaging, which is quicker than electron microscopy and more able to deal with impurities, but can still see objects that are 500 times smaller than the width of a human hair. First, Laine et al. took pictures of many viruses with this new imaging technique, sorted the images into groups based on their appearance, and then trained computer algorithms with the pre-sorted groups (a technique called “supervised learning”). Next, the trained algorithms were shown new images of viruses and asked to classify them. The algorithms could separate images of a mixed population of viruses into six groups according to their shape and size, and then analyse each group in a specific way. For example, they would measure and report the length of filament-shaped viruses, the radius of spherical viruses and the length and width of rod-shaped viruses. The first set of test images were of Newcastle Disease Virus, which is currently under development as a treatment for cancer. But further testing revealed that the algorithm also works for the influenza virus, which is used to make flu vaccines. The algorithm could classify the viruses in pure and impure samples, and the imaging technique could handle over 200 viruses each second. This approach of combining super-resolution imaging with artificial intelligence could help scientists to understand what makes good vaccines and how best to optimise the production of viruses for medical purposes. It could also allow researchers to respond more rapidly to outbreaks of viral infections. The next step is to build this work into a system that can be used by the pharmaceutical industry.