Here are some of the projects using AI to address the coronavirus outbreak:
AI in Drug Discovery
A number of research projects are using AI to identify drugs that were developed to fight other diseases but which could now be repurposed to take on coronavirus. By studying the molecular setup of existing drugs with AI, companies want to identify which ones might disrupt the way COVID-19 works.
BenevolentAI, a London-based drug-discovery company, began turning its attentions towards the coronavirus problem in late January. The company's AI-powered knowledge graph can digest large volumes of scientific literature and biomedical research to find links between the genetic and biological properties of diseases and the composition and action of drugs.
The company had previously been focused on chronic disease,
rather than infections, but was able to retool the system to work on COVID-19
by feeding it the latest research on the virus. "Because of the amount of
data that's being produced about COVID-19 and the capabilities we have in being
able to machine-read large amounts of documents at scale, we were able to adapt
[the knowledge graph] so to take into account the kinds of concepts that are
more important in biology, as well as the latest information about COVID-19
itself," says Olly Oechsle, lead software engineer at BenevolentAI.
While a large body of biomedical research has built up around chronic diseases over decades, COVID-19 only has a few months' worth of studies attached to it. But researchers can use the information that they have to track down other viruses with similar elements, see how they function, and then work out which drugs could be used to inhibit the virus.
"The infection process of COVID-19 was identified relatively early on. It was found that the virus binds to a particular protein on the surface of cells called ACE2. And what we could with do with our knowledge graph is to look at the processes surrounding that entry of the virus and its replication, rather than anything specific in COVID-19 itself. That allows us to look back a lot more at the literature that concerns different coronaviruses, including SARS, etc. and all of the kinds of biology that goes on in that process of viruses being taken in cells," Oechsle says.
The system suggested a number of compounds that could potentially have an effect on COVID-19 including, most promisingly, a drug called Baricitinib. The drug is already licensed to treat rheumatoid arthritis. The properties of Baricitinib mean that it could potentially slow down the process of the virus being taken up into cells and reduce its ability to infect lung cells. More research and human trials will be needed to see whether the drug has the effects AI predicts.
Shedding light on the structure of COVID-19
Human epidemiologists at ProMed, an
infectious-disease-reporting group, published their own alert just half an hour
after HealthMap, and Brownstein also acknowledged the importance of human
virologists in studying the spread of the outbreak.
"What we quickly realised was that as much it's easy to
scrape the web to create a really detailed line list of cases around the world,
you need an army of people, it can't just be done through machine learning and
webscraping," he said. HealthMap also drew on the expertise of researchers
from universities across the world, using "official and unofficial
sources" to feed into the line
list.
The data generated by HealthMap has been made public, to be
combed through by scientists and researchers looking for links between the
disease and certain populations, as well as containment measures. The data has
already been combined with data on human movements, gleaned from Baidu, to
see how population mobility and control measures affected the spread
of the virus in China.
HealthMap has continued to track the spread of coronavirus
throughout the outbreak, visualising its spread across the world by time and
location.
Spotting signs of a COVID-19 infection in medical images
Canadian startup DarwinAI has developed a neural network
that can screen X-rays for signs of COVID-19 infection. While using swabs from
patients is the default for testing for coronavirus, analysing chest X-rays
could offer an alternative to hospitals that don't have enough staff or testing
kits to process all their patients quickly.
DarwinAI released COVID-Net as an open-source system, and
"the response has just been overwhelming", says DarwinAI CEO Sheldon
Fernandez. More datasets of X-rays were contributed to train the system, which
has now learnt from over 17,000 images, while researchers from Indonesia,
Turkey, India and other countries are all now working on COVID-19. "Once
you put it out there, you have 100 eyes on it very quickly, and they'll very
quickly give you some low-hanging fruit on ways to make it better,"
Fernandez said.
The company is now working on turning COVID-Net from a
technical implementation to a system that can be used by healthcare workers.
It's also now developing a neural network for risk-stratifying patients that
have contracted COVID-19 as a way of separating those with the virus who might
be better suited to recovering at home in self-isolation, and those who would
be better coming into hospital.
Monitoring how the virus and lockdown is affecting mental
health
Johannes Eichstaedt, assistant professor in Stanford
University's department of psychology, has been examining Twitter posts to
estimate how COVID-19, and the changes that it's brought to the way we live our
lives, is affecting our mental health.
Using AI-driven text analysis, Eichstaedt queried over two
million tweets hashtagged with COVID-related terms during February and March,
and combined it with other datasets on relevant factors including the number of
cases, deaths, demographics and more, to illuminate the virus' effects on
mental health.
The analysis showed that much of the COVID-19-related chat
in urban areas was centred on adapting to living with, and preventing the
spread of, the infection. Rural areas discussed adapting far less, which the
psychologist attributed to the relative prevalence of the disease in urban
areas compared to rural, meaning those in the country have had less exposure to
the disease and its consequences.
There are also differences in how the young and old are
discussing COVID-19. "In older counties across the US, there's talk about
Trump and the economic impact, whereas in young counties, it's much more
problem-focused coping; the one language cluster that stand out there is that
in counties that are younger, people talk about washing their hands,"
Eichstaedt said.
"We really need to measure the wellbeing impact of
COVID-19, and we very quickly need to think about scalable mental healthcare
and now is the time to mobilise resources to make that happen," Eichstaedt
told the Stanford virtual conference.
Forecasting how coronavirus cases and deaths will spread
across cities – and why
Google-owned machine-learning community Kaggle is setting a
number of COVID-19-related challenges to its members, including forecasting
the number of cases and fatalities by city as a way of identifying
exactly why some places are hit worse than others.
"The goal here isn't to build another epidemiological
model… there are lots of good epidemiological models out there. Actually, the
reason we have launched this challenge is to encourage our community to play
with the data and try and pick apart the factors that are driving difference in
transmission rates across cities," Kaggle's CEO Anthony Goldbloom told the
Stanford conference.
Currently, the community is working on a dataset of
infections in 163 countries from two months of this year to develop models and
interrogate the data for factors that predict spread.
Most of the community's models have been producing
feature-importance plots to show which elements may be contributing to the
differences in cases and fatalities. So far, said Goldbloom, latitude and
longitude are showing up as having a bearing on COVID-19 spread. The next
generation of machine-learning-driven feature-importance plots will tease out the
real reasons for geographical variances.
"It's not the country that is the reason that
transmission rates are different in different countries; rather, it's the
policies in that country, or it's the cultural norms around hugging and
kissing, or it's the temperature. We expect that as people iterate on their
models, they'll bring in more granular datasets and we'll start to see these
variable-importance plots becoming much more interesting and starting to pick
apart the most important factors driving differences in transmission rates
across different cities. This is one to watch," Goldbloom added.
~ Jai Krishna Ponnappan
~ Jai Krishna Ponnappan
