Special Feature
Part of a ZDNet Special Feature: Coronavirus: Business and technology in a pandemic

How a smartphone coupled with machine learning may become a simple, efficient test for COVID-19

A scientist at the University of Cincinnati has been refining an ingenious design for an ultra-portable genetic test that uses AI and smartphones. It could eventually offer a cheaper, simpler alternative for detecting coronavirus infections.

The smartphone may be the new frontier in COVID-19 testing

"Testing in the United States is a mess" wrote the editors of Nature magazine recently, addressing a series of missteps that have led to a serious shortfall in the country's ability to detect COVID-19

"With shambolic coordination at the national level, existing providers have become swamped, testing backlogs have lengthened and shortages of swabs, supplies and reagents have grown," observed the editors. 

latest developments

Coronavirus: Business and technology in a pandemic

From cancelled conferences to disrupted supply chains, not a corner of the global economy is immune to the spread of COVID-19.

Read More

The disaster of US testing for COVID-19 opens the way for novel approaches to diagnosing the disease, some of which may get emergency-use approval from the US Food & Drug Administration if they show promise.

Enter the smartphone. The camera on the average smartphone is powerful enough to capture substantial information from the appearance of fluid in a genetic test. Coupled with deep learning computer programs, such detail could, in theory, be analyzed by an app on the phone to produce an answer on the spot as to whether an individual has a disease. 

A smartphone approach could open the way to a very simple diagnostic kit for COVID-19 that could be used on-site, whether in a clinic, a drive-through or pop-up testing facility, or an out-of-the way setting such as a rural location that's far from a laboratory. 

The science of such an approach has been in development for several years now and is reaching a new level of elegant simplicity in the time of COVID-19. 

Aashish Priye, an assistant professor in the Department of Chemical and Environmental Engineering at the University of Cincinnati, has been steadily refining an ultra-portable test-in-a-box, if you will, for infectious disease.  

With just a heating plate, an LED light source, and a smartphone, Priye's design for a portable, battery-operated kit can be used to test for the presence of virus in under an hour. In the latest version, a deep learning neural network known as a convolutional neural network, or ConvNet, can be used to tell how much viral material is in a sample of blood or saliva, based on the pixels captured by the smartphone camera.

priye-et-al-2017-lamp-box-materials.png

In a 2017 research paper in Scientific Reports of Springer Nature, Aashish Priye and colleagues described a "LAMP box" test kit, a radically simplified genetic test that used a smartphone to detect the presence of viral RNA in solution. Priye is now working on an even simpler approach that uses the smartphone to measure the cloudiness of solution, and therefore the viral RNA content, via a machine learning approach known as convolutional neural networks.

Priye et al. 2017

"We are exploring nontraditional image analysis techniques to improve the detection capabilities of consumer smartphones in order to perform virus detection using something extremely powerful but at the same time widely accessible to most people," Priye explained in an email exchange with ZDNet

Priye's work is itself made possible by the radical simplification of genetic testing in the past two decades, an increasingly popular approach known as LAMP. 

All nucleic acid-based diagnostic tests, including LAMP, are based on a controlled chemical reaction, the result of which is what's known as amplification. Put some DNA, the genetic instructions of life, or, in the case of a virus, RNA, the single-stranded working copy, in a polymer tube, along with some short strands of nucleic acid, known as primers, and an enzyme known as polymerase; heat and cool the mixture repeatedly; and the result is that the DNA or RNA will become multiplied several-fold.

Also: 'We are doing in a few months what would normally take a drug development process years to do': DoE's Argonne Labs battles COVID-19 with AI

The process was first described in 1985 by scientists at Cetus Corporation, the San Francisco biotech startup, who called it the "polymerase chain reaction," or PCR. Kary Mullis, the principle inventor at Cetus, won the Nobel Prize for the procedure. PCR revolutionized biology by making it simple to create massive amounts of DNA or RNA for study.

Mullis's insight back at the beginning was that PCR doesn't just amplify, it also picks out specific genetic sequences of nucleic acid by extracting them from the rest of the surrounding genetic material. That means DNA or RNA that's not yet fully known can be identified in solution, a fact Mullis noted was "as important as the amplification itself." Amplification is not just multiplying DNA and RNA, in other words, it's a tool for identification and therefore a tool for diagnosis.

PCR has become ubiquitous in labs, and a version, quantitative reverse-transcriptase PCR, or RT-qPCR, is the gold standard for COVID-19 tests. But there are several cumbersome aspects of PCR that weigh down its use, especially during a pandemic when speed of testing is of the essence. 

The full process of PCR typically takes several hours to perform, plus the time needed to ship samples to the lab where the PCR equipment resides. The main piece of equipment, a thermocycler, which carefully guides the chemical mixture through the repeated stages of heating and cooling, costs thousands of dollars and requires costly reagents such as the Taq polymerase at the heart of the process. That puts it beyond the reach of anyone but well-funded labs.

priye-et-al-2018-color-space-of-virus.png

In a follow-on paper in 2018, Priye and colleagues refined the color-space use of the smartphone camera to quantify the fluorescent reaction of viral RNA in LAMP, including the Zika virus.

Priye et al. 2018

"Right now the way for testing in most cases is someone comes, or you have to send your sample to a place where there's a PCR machine, and you do PCR, and then you get back the results, and turnaround time is hours to days," Priye explained to ZDNet. "You want to get that down to less than an hour, maybe 30 minutes."

Attempts have been made to simplify PCR. For example, a number of companies are working on portable versions of PCR, a trend referred to as DIY Bio. But the state of the art in cheap amplification approaches is the alternative method known as LAMP, which stands for loop-mediated isothermal amplification

Introduced to the world in 2000 by scientists from Eiken Chemical of Japan, LAMP can be performed in under an hour. Just as important, the sample only needs to be heated to one temperature rather than the multiple ups and downs of heating and cooling in PCR. The simplicity of the single temperature, or isothermal, preparation makes equipment for LAMP an order of magnitude cheaper than PCR thermocyclers.

After amplification, the DNA or RNA, if it's been found, needs to be detected, and that is where Priye has made major strides. 

"The uniqueness in our lab is combining the portability aspect of LAMP with image analysis on a smartphone app to detect LAMP products using the inbuilt camera," he explained.

The standard way to read the results of LAMP is to hit the sample with light that makes the detected DNA or RNA fluoresce and then observing the colors in that fluorescence. A light emitting diode, or LED, can be used as the light source and typically a device called a fluorometer measures the color response. 

In a 2017 paper, Priye and scientists at Sandia National Laboratories proposed a "LAMP box" test kit with a heating pad, an LED light, and a smartphone in place of the fluorometer. He was able to remove a piece of equipment, thereby simplifying the overall design.

In his present work, Priye is simplifying things even further by using what he calls a "label-free" detection method that doesn't need to incorporate fluorescent molecules.

"Even though fluorescently labeled primers are effective, they are expensive to implement in LAMP assays, especially when optimizing different primer sets," he said.

It turns out there is a much simpler way to see a positive reaction in LAMP.  In 2001, shortly after they introduced the technique, the Eiken scientists described an interesting fact. A positive result in a test naturally produced a by-product, an opaque, milky white substance within the liquid solution. That turbidity, as it's called, is magnesium pyrophosphate, which is formed as a result of the interaction of ions in the LAMP process. The substance can be seen with the naked eye just by looking at the sample in the tube after the amplification step. 

Although you can see turbidity directly, it's important with diagnostic tests to be able to measure just how much turbidity there is, because it can tell you how much of the target DNA or RNA is present in a sample. 

That problem of measurement is where machine learning comes in. In collaboration with a colleague, Surya Prasath, Priye is working on quantifying the turbidity by analyzing high-resolution images with a neural network.

Also: Scientists led by AI scholar Yoshua Bengio urge privacy, opt-in approach for COVID-19 contact tracing

"We plan to train and test a deep convolutional neural network to classify 10-times magnified images according to the level of precipitation formed in the LAMP samples," explained Priye. ConvNets, as they're known, have been the workhorse of modern deep learning for a range of applications but especially for classifying and categorizing images. 

Priye and Prasath are also testing simpler forms of machine learning in parallel, he told ZDNet, including support vector machines and random forests, to determine the optimal model to use. The experiments involve both laboratory-generated images, as well as synthetic images generated by a generative adversarial network, a kind of deep learning neural network that can synthesize images based on samples of the original. 

Once the optimal neural network is developed, it will be incorporated in a detection app running on the smartphone, said Priye.

The heating element is another part of the kit Priye is working to simplify. In the design outlined in 2017, a heating plate is powered via a five-volt battery. He has been experimenting with a light-emitting diode at about 1 watt and 1 amp of driving current that can heat small amounts of fluid resting on a thin film of metal. The process is called photonic heating, converting light into heat. 

The outlines of an extremely simple kit start to come into view: no fluorescence system, no heating pad, just a thin metal plate, some plastic tubes, a light source and a smartphone. 

Priye hopes to have a prototype of the kit by the end of the year. At the same time he is pursuing the diagnostic equipment, he is moving on a second front to simplify LAMP itself. There is one area where LAMP is more cumbersome than PCR, and that's in the development of the primers, the strings of nucleic acids used to bracket target DNA or RNA to be amplified. PCR only needs two primers to detect each viral infection, whereas LAMP requires 6 to 8 primers. 

More primers complicates what can happen during amplification.

"With a greater number of primers, there is a greater chance of these primers interacting with each other and the wrong part of the virus RNA," which can lead to skewed results, explained Priye. "This is one of the major reasons behind these reactions showing false positives."

Typically, the design of primers for LAMP, which has to be carried out with each new disease, takes months, with graduate students in the lab handling the grunt work of trial and error in search of the right recipe.

To help streamline that work, Priye is constructing software based on thermodynamical models, what he calls a "LAMP reaction simulator," which can capture the "intermediate reaction products" in LAMP -- it can model what's likely to happen during amplification.

"By predicting the likelihood of these primers to interact with each other, before even going into the laboratory, we can greatly reduce the time required to design and test multiple primers for a particular virus," explained Priye. 

It's rather like rapid prototyping or digital modeling for manufacturing, where attempts can be iterated before committing to a final design.

The timeframe for completing that software may be sometime next year, said Priye.

Priye's stripped-down test kit and simulation software is part of LAMP's evolution into a platform. Efforts are popping up all over to build upon the simple amplification method. 

For example, startup Mammoth Biosciences, co-founded by U.C Berkeley scientist Jennifer Doudna, who pioneered the use of CRISPR/Cas gene editing, has developed a version of LAMP that uses the Cas12 enzyme to detect the results of LAMP. The enzyme automatically cuts genetic material when the target DNA or RNA shows up in the amplification step. The test can be performed in a half an hour, and the detection step makes use of simple strips called lateral flow immunoassays, similar to a pregnancy test.

Such ingenious mash-ups of lightweight approaches suggest COVID-19 is resulting in a wave of innovation that will change the nature of testing and diagnosis for years to come. Maybe they won't immediately resolve the testing bottleneck in the case of COVID-19, but there is the possibility that the present debacle, as bad as it is, will have a tiny silver lining for future generations.