How Machines Have Revolutionized Medicine | WIRED Brand Lab

(whimsical music)

[Narrator] The scientific method has been the bedrock

for how we learn about the world around

us ever since the 17th century.

Now, machine learning is helping to change that,

thanks in part to researchers working

in drug discovery at Novartis.

(whimsical music)

Getting drugs to market has traditionally been a long road.

It takes about 12 to 15 years

to go through several million molecules,

and bring just one of them to the patient at the end.

Anything that can reduce the time

or cost on that would be a win.

[Narrator] One of the most time-intensive tasks

in this process is image analysis.

You would have to look at row

by row results of experiments.

We would have a chemist or a biologist poring

over these for months.

While the researchers bring their

own experience to the table, they're

also bringing their own biases,

and can possibly be limiting themselves

by only having certain hypotheses

that they're using to interpret the data,

as opposed to ones that are just

simply outside their domain.

[Narrator] Enter the deep neural network.

(electronic music)

A deep neural network is a sophisticated type

of machine learning model.

It works by finding patterns in data,

and this is done by breaking apart the data

into different features.

It's like if you imagine you're trying

to describe what a cat looks like.

One layer of this network might be the overall shape

of cats from photos.

Another layer may be more about the ear

or the whisker, and then the output is

more of a wholistic view of what a cat looks like.

We are currently using deep neural networks

in order to predict how compounds work

in cells, and the way we do this is

by collecting images of cells using a microscope.

We dye the cell, which enables

us to create a beautiful image

of what the compound has done to the cell.

Now, at this point, you have a collection

of images, and that is the input

into the deep neural network.

Machine learning is able to do,

just in minutes, what it would take a chemist

or a biologist months.

It can find patterns much more quickly, and build models

that would make a prediction much more accurately.

The hope is to essentially separate the wheat

from the chaff, of molecules

that will work versus the molecules that don't work,

and if we can filter out the molecules

that would have otherwise have failed,

further down the pipeline, then

we can speed the development of getting drugs

to the patient, and also reduce costs, as well.

Machines are unlikely

to replace scientists in drug discovery.

They are great at solving specific tasks

with lots of data.

There's other kinds of tasks in drug discovery

that involve more integrative, creative thinking.

They're now free to spend more

of their time on the interpretation

of the results, which is something machine learning can't

necessarily do, and other tasks

that just require more of their attention and time.

(synthesized music)

[Narrator] With pioneers like those

at Novartis, harnessing the power

of machine learning has the potential

not only to save lives, but to change science as we know it.

What we're doing is a paradigm shift

from the traditional scientific method.

Where originally you would have observations

that lead to a specific hypothesis,

and then you test that hypothesis,

in this case, we're generating large-scale data

in an unbiased way, and then we're turning algorithms

in machine learning onto the data

to find new hypotheses that we can then go test.

We are creating a more successful future

for understanding and interpreting experiments.

It's basically a meta-analysis approach to science.