We’re Teaching Robots and AI to Design New Drugs

[♪ intro]

Every year, there are thousands of new drugs  in development at labs around the world.  

But only a tiny fraction of them make it through  to human trials -- never mind final approval.

And even among those that do get through, the  majority of new medications are actually just  

new-er versions of existing drugs,  like cheaper generic versions.

But there’s a new hero in town with  the potential to help things along:  

drug discovery systems based  on artificial intelligence.

Even though the tech is still growing,  

several new drugs designed with the aid  of AI are already in clinical trials.

And together with other modern tech  like fully automated robot scientists,  

AI is changing drug discovery in  revolutionary ways -- and it’s  

accelerating the process beyond  anything we’ve seen before.

It’s only recently that we’ve been  able to create drugs from scratch.

Some drugs are based on traditional remedies  and natural products -- like aspirin. 

The first fully synthetic drug,  the sedative chloral hydrate,  

was developed in 1869. But there was still  quite a bit of trial and error involved. 

Rather than throwing spaghetti  at the wall to see what sticks,  

it was… throwing chemicals at people  to see if they stop being sick.

That’s how, in 1932, we got sulfa drugs,  when researchers at a chemical dye company  

discovered some of those dyes could  be used to kill dangerous microbes. 

And pure random luck also did a lot for the  early pharmaceutical industry. Most famously,  

if Alexander Fleming had been a little more  fastidious about keeping his lab clean,  

we might not have penicillin.  

But all that has changed. Starting in  the latter half of the 20th century,  

we’ve seen the rise of rational drug design.

That means no more spaghetti-throwing. Instead,  

scientists build drugs from the ground up,  driven by hypotheses of how they might work. 

Having so much data to rely on can  be a double-edged sword, though.

In fact, in a sense, you could say we’ve gotten  too good at this. The number of researchers and  

person-hours it takes to investigate every  single drug candidate is truly monumental.

But now, artificial intelligence is turning  the tables by helping to guide and accelerate  

this process. It’s helping identify drug  candidates we might not have found otherwise,  

and cutting years off their development.

AI systems can be given a general  description of what we want to find,  

analyze a bunch of literature and databases,  and pick out the best hits for later research.

This starts right at the beginning: before you  even think about how your new drug would work,  

you have to decide what to  design it for. You need a target.

That target is one of the dominoes in the sequence  

of things that happen in your  body to create a disease.

It could be a mutated gene, or an enzyme  that’s working harder than it should. 

There’s usually a huge body of scientific  literature devoted to the disease you want  

to treat -- stuff like research articles,  clinical trial reports, and patient records. 

Deciding on a target based on all this information  is like looking for a needle in a haystack,  

and it’s a lot to ask of our brains.

AI can help us see the big picture -- then pick  

out the one really specific thing  we needed from that big picture. 

AI systems use a bunch of  different technologies to do that,  

but generally, the key is  natural language processing.  

Natural language processing is basically  what allows your phone’s voice assistant  

to understand you when you say  “Play something by Tchaikovsky.” 

If you’re lucky, it will understand that you  want to hear any music composed by Tchaikovsky,  

instead of a specific piece  he wrote called “Something.” 

Now, the AI doesn’t really  understand what a “Tchaikovsky” is.   

But by analyzing the relationships between  the words and the context, it can tell that  

you want to look for “Tchaikovsky” in the part  of its database that holds a list of composers.

And the principle is the same with drug discovery.

So even though the AI doesn’t understand what  a “gene” is, if you instruct it to sort through  

the literature on a given disease, it  can identify the ones that stand out.

It can also find the places in the literature  that talk about that gene -- and look at the  

words connected with it. Do authors refer  to it often in the context of that disease?  

And crucially, do they talk  about a causal relationship?

If so, the AI can conclude  that that gene is worth a look.

Natural language processing also means that an AI  

system doesn’t need this info to be  pre-formatted by an army of people. 

You can feed it things written by a human, like  scientific articles or medical case reports. 

So AI-based systems can analyze enormous  amounts of data and give us meaningful answers. 

And they’re much faster readers than we are,  so they can go through way more information,  

and handle way more complexity, than humans alone.  Plus, they never need to duck out for coffee.

This vastly speeds up the process  of finding potential drug targets  

that we might have found eventually,  but it also makes it possible to find  

completely new targets that  we might not have noticed. 

For example, researchers at the US company Berg  

grew cancerous and healthy cells from  over a thousand donors in petri dishes. 

They varied the conditions and tracked  the unbelievably complex chemistry that  

results from cells simply doing their thing.

They ended up with trillions of  pieces of data from those samples.

That’s a lot of raw info about what happens in  a cancer cell compared to a healthy one -- which  

is to say, too much. The amount of data  they had was so gigantic that without AI,  

it would have effectively been useless.

But their AI system was able to analyze  this data and identify various molecules  

that were out of whack in the cancer  cells compared to the healthy ones. 

This suggested a new cancer drug target  – a molecule called coenzyme Q10. 

And a new drug candidate they  developed based on that AI discovery  

is now in clinical trials for pancreatic  cancer and squamous cell carcinoma. 

So now we know what we want to target with a  novel drug. Which means we need… a novel drug. 

Meaning it’s time to identify and synthesize  

chemical compounds that will hopefully  hit that target in just the way we want.

But how do you choose what to synthesize?

Even before AI helpers, researchers were able  to make educated guesses about what would work,  

by looking at the chemical  features of substances that  

they already knew could interact with the target.

That interaction doesn’t have to be beneficial  – if something sticks to your target,  

that gives us a clue for how to design  something new that will stick to it as well. 

But still, even if researchers know what chemical  features they want their drug candidate to have,  

that may still mean thousands of options.

And AI systems come to the rescue here as well.

Scientists can tell the AI what  chemical parameters they’re looking for,  

and the system will not only  find suitable candidates,  

but cut the list down to those  that will potentially work best. 

The process is kind of similar to things  like using AI for image recognition.

When you do an image search for “cats,”  

you may get 99 images of felines, and  one gorgeous kitten-looking cloud.

Because of those occasional blunders, the  AI doesn’t actually make the decisions.

Scientists still evaluate the results,  

and the AI system only helps automate  and accelerate the grunt work. 

But, like… that’s really helpful.

For example, by using AI to help choose  chemical compounds, the makers of a new  

candidate drug for obsessive-compulsive disorder  were able to cut down their development cycle  

from around five years to one, and get  their candidate into clinical trials.  

Once you’ve decided on a drug candidate  

and synthesized the right  chemical, it’s time for testing.

But before human tests, and before animal  tests, researchers start with simpler assays. 

That means testing the chemical  using cultured cells or cell-free  

cocktails containing the drug’s potential target.

It used to take years to do this preliminary  testing for a single potential medication. 

But in the 21st century, pharmaceutical companies  have turned to robotic high-throughput screening,  

which makes it possible to test hundreds  of thousands of compounds in a single day. 

A human might have to pipet the hundreds or  thousands of candidate compounds into one  

cell culture dish at a time, but a robot  can quickly zip through a bunch of them. 

A researcher can just design the  experiment and then check on the results.

Like the Voltron of modern pharmaceutical science,  

a bunch of autonomous capabilities can also be  combined into what’s called a robot scientist. 

That system uses AI to identify specific  experiments with a lot of potential,  

and can then autonomously use lab  equipment to perform these experiments  

and fine-tune its decisions of where  to go next based on their results. 

For example, Eve, an AI-equipped robot  scientist at the University of Cambridge,  

has already identified a new  potential treatment for malaria.

Eve first identified a list of  compounds that may counteract malaria,  

then screened them against yeast cells in culture  to see which of those chemicals worked best.

This way, the Cambridge team was able  to report that Eve had identified the  

well-known antimicrobial compound triclosan  

as a candidate to help combat  treatment-resistant strains of malaria. 

Even with all of this futuristic tech, though,  

only a few dozen new drugs  are approved every year.

For example, in 2019, the US Food and Drug  Administration approved 48, but only 20 of them  

were actually distinct enough from existing  medications to be considered truly new. 

That’s because discovering a new drug and  bringing it to market is an expensive,  

long and difficult process.

To develop one new drug candidate, researchers  need to screen up to 10,000 compounds.

And on average, only five of those turn out to  be good enough to go into clinical testing. 

If it gets that far, a drug candidate needs  to complete three phases of clinical trials.  

Of those, 90%  fail to gain FDA approval. 

And that’s why developing a new  drug may take up to 15 years,  

and cost around 1.3 billion dollars. Each. 

But smart AI systems like the ones we’ve talked  about today are in a position to change that.

We’re only beginning to see the results  of this AI revolution in drug discovery. 

There are literally hundreds of companies  

developing AI-based systems for  the pharmaceutical industry. 

Human scientists still drive the process. And  while machines are great at finding patterns and  

sorting through a lot of information in a hurry,  our brains can do a whole lot that machines can’t.

But the help that AI represents  is already truly revolutionary.

It can cut the time and cost  necessary to develop a drug.  

And by doing so, it’s widening that pipeline  of potentially life-saving treatments. 

When candidate drugs inevitably flunk out  -- because they don’t work, or aren’t safe  

for us to use -- this technology ensures  there will be more waiting for us to test.

And that’s great for all of us -- even if it means  

the future looks kind of like a  robot handing us a miracle pill.

We aren’t robots, and we don’t  have any miracle drugs handy,  

but what we do have are  pins for sale over at DFTBA.

Specifically, today is your very last chance  to grab this month’s Pin of the Month -- which  

is of the Chandra X-ray Observatory. It’s  a very charming looking little telescope,  

even if it kind of looks like a  traffic cone laying on its side…

Either way, this pin will only be on sale  through midnight tonight, September 30th.  

After that, we’ll have another pin available  for pre-order and start shipping this one out!  

To get your hands on it, check  out the link in the description.