The focus of this programme is on understanding the disease risks posed by entomopathogens on the health of insects that are currently reared en masse in industry. However, gaining an understanding of the natural occurrence of these disease in the wild could also provide useful information, not just in the insects of rearing interest but also for the majority of the ones that are not.
In one of my projects, I took advantage of some great technology a colleague was using. This piece of kit is an optical insect sensor, which allows for automated monitoring of insect flights, and is used remotely. The way it work is that it emits near-infrared light and when an insects flies in front of it, the backscattered light reflected off the insect is picked up by the sensor. Using this, measurement such as wing beat frequency, number of flight events and object size can be used to determine the insect species flying in past. Talk with my colleague got us thinking, can this technology be used to detect whether or not an insect is diseased?
The E. muscae-housefly system is a great model to test this hypothesis on, as the infection period is very fixed. Once a fly become infected, six or seven days after, the fly is behaviourally manipulated by the fungus, four to six hours before sunset. After being manipulated, the fly is ‘killed’ and the fungus eats the remaining nutrients inside the fly before bursting out the skin, shooting spores around and releasing pheromone-like chemical to infect the next fly. Before this dramatic end of life event, the fungus grows in number throughout the week-long infection.
Using two cages large enough so for flies to fly around, we release healthy male flies in one and flies we infected into the other and left them record data for a whole week.
After training a machine learning (ML) algorithm on some of the data, we used that model to analyse the rest of the data to see if we could determine if a fly was ill or not. Comparing each day after infection to the same day for uninfected, we see that the accuracy of the model prediction, i.e. whether the fly is healthy or infected, increased significantly. With flies infected for 2 days, the model could determining insect health with 61% accuracy. Moreover, after 5 days, the accuracy rose to 99%! This shows that ML in combination with near-infrared sensors can be used to distinguish insect health. This work could be used for directly monitoring certain reared insects maybe, but also mosquito-borne human diseases.
If this work interests you, check out the preprint of this article here:
E. Bick, S. Edwards, H. H. De Fine Licht (2021) Detection of insect health with deep learning on near-infrared sensor data, bioXriv. https://www.biorxiv.org/content/10.1101/2021.11.15.468635v1.full.pdf+html