2018-12-01_Discover

Body mass

Under1kg?

Geographic

range

Under 150,000

km^2?

TREE #2 TREE #3

TREE #1

Life span

Less than

1 year?

Age of

sexual maturity

More than

50 days?

Group

population

size

Less than

10?

Litters

per year

Less than 3?

Period of

activity

Active during

day?

Ye s

Ye s

Ye s

No

No

No Yes No Yes No Yes No

Ye s N o

58% identied correctly

83% identied correctly

rp

oc

es

s^ r

ep

ea

ts

67% identied correctly process repeats

non-disease carrier disease carrier

unlabeled for algorithm

carrier non-carrier carrier non-carrier carrier non-carrier carrier non-carrier

Misclassiied

Not known to transmit disease

Known to transmit disease

Not yet classiied by algorithm

Key

60 DISCOVERMAGAZINE.COM

TOP: COURTESY OF BARBARA HAN. BOTTOM: ALISON MACKEY/DISCOVER. SHUTTERSTOCK ELEMENTS FROM BASEL101658, POTAPOV ALEXANDER, HEIN NOUWENS, A SK (CREATURES); BLACK CREATOR (MAN)

A Learning Process

In the world of epidemiology, diseases

that have seen an uptick in recent

years are called “emerging infectious

diseases.” But are there really more

cases of these diseases, or have we

just become better at spotting them?

According to Barbara Han, a disease

ecologist at the non-proit Cary

Institute of Ecosystem Studies in New

York, it’s not just us getting better.

“It’s actually an increasing problem of

infectious diseases,” she says. And most

of these diseases originate in animals.

Han decided to igure out what

makes certain animals more likely

to host speciic diseases. “There is

something inherent about a species that

enables it to carry disease, compared to

the vast majority that don’t,” she

says. “I want to know what the

data can give me, what can the

data show me, about what distin-

guishes those two.” She turned to

algorithms and machine learning.

Han starts with a list of species

that researchers have already

agged as disease carriers or

non-disease carriers. She then

trains a computer algorithm

to separate the species on the

list — not labeled in any way, so the

algorithm doesn’t know which is which

— by dozens of traits. For example,

the algorithm may start by looking at

an animal’s body mass, followed by its

age of sexual maturity and inally by

whether it’s nocturnal or not.

At the end of this sorting, the

algorithm will ideally have

grouped species by whether

they’re disease carriers or not.

But this irst sort gets a fair

bit wrong. To make the algo-

rithm more accurate, Han has

the computer do another round

of sorting, this time focusing

on the species it miscategorized

the irst time. When it does

this over and over again, the algorithm

learns. And, importantly, it learns which

factors contribute to a species carrying a

transferable disease or not. “At the end

of that process, you get a very powerful

predictor,” Han says. When the model

To train an algorithm to identify

zoonotic species, known disease

carriers and non-disease carriers

from an animal group, in this case

rodents, are fed into the algorithm

as unlabeled data points. NOTE: Only

around 10 percent of disease carriers are

currently known, but we show a 50/50

split here, for simplicity.

The algorithm then sorts them into

groups using randomly selected traits.

The algorithm gets a

lot wrong (hollow

boxes) in its first

pass, so it repeats

the process multiple

times with other

randomly selected

traits, focusing on

misclassified species.

With each attempt, the

algorithm learns which

traits are most likely to

crop up in disease carriers.

Diseases that can

pass between

animals and humans

are called zoonotic.

1

2

3

Barbara

Han

disease

ecologist,

Cary Institute

of Ecosystem

Studies