Continuous Monitoring of Animal Health using Computer Vision and Deep Learning

TRKIR — Detailed Phenotypic Data on Animal Models of Human Disease

A tremendous amount of data is being collected from cell lines, animal models and patients to understand disease and be able to prevent or treat it. A major part of this discovery has relied on accurate data that can be generated with controlled experiments and yet preserve the characteristics of “real-world” human disease. This challenge has led to the emergence of humanized animal models, bioreactors for cell lines and organoids. However, even with these new technologies, sampling the data in a way that allows us to go “deep” into molecular detail at single cell resolution has not been met with equally detailed phenotypic data capture.

Our team first met this challenge working on the DARPA INTERCEPT project that involved a new type of a vaccine, designed to co-evolve with the natural virus while “stealing” resources to slow down replication. This amazing idea has led to a number of labs across the US to develop defective genomes (or particles) that can be rapidly produced to interfere with natural infection to slow down a sudden outbreak. And even though this technology was not ready for the current pandemic, many teams did show significant progress in this ambitious program.

At the same time, everyone knew that such a treatment, even if it would be possible, would need to be tested extensively to understand dosages, timing and route of intervention that does not cause more harm than good. That is where an animal model that replicates human disease characteristics can be invaluable at generating detailed data on disease progression and dynamics.

As a result, our team developed a novel phenotyping system to capture highly detailed data on animal physiology and behavior, allowing individualized and long-term collection of data. This system enables remote tracking of changes, condition assessment based on peronal controls and near-real time anomaly detection.

Manifestations of disease are gradual and clinical symptoms are evident only after severe changes have already taken place internally. To overcome this challenge, we knew that the system has to be sensitive enough to pick up minor changes that evade the human eye or short-term testing. We discovered that in many cases, this happens not in a clear physiological sense, but at first as interruptions of 12, 3 and 1-hour cycles in activity. This discovery solidified our appreciation for the long-term monitoring coupled with analysis of periodicity (or “seasonality decomposition”).

Soon enough, we had a working prototype — reliably and accurately collecting data on several animals at a time 24/7 for days or even weeks at a time. Once the basic hardware works without the hassle, the ideas of new experiments, data and analysis start coming. We realized that this is not just a solution for our own project, but a platform that others can leverage to collect detailed data of various kinds in experiments we have not even dreamed of.

Now, the system is being piloted, helping us better understand how different environments, animal models and even labs affect the data. The road ahead will be full of surprises — I am sure that as the technology matures, it will be used to identify plenty of new insights that we will all benefit from.

You can learn more at our website: https://pheno.trkir.com

or view this brief video: https://www.youtube.com/watch?v=HXb_oVgAPq8