If you’re at all familiar with the biotech space, you’ve probably heard of the promises of artificial intelligence and machine learning.

These promises have been tantalizing. AI and ML can, proponents claim, help researchers speed along the arduous and expensive drug development process. By attempting to utilize algorithms and computing power so dynamic as to essentially eliminate menial tasks and trial and error altogether, the technology comes with big expectations likening it to a new industrial revolution.

It’s gotten to the point where the AI space, be it intentional or in some self-fulfilling prophecy, has cultivated the notion that the possibility of creating new drugs at the push of a button could be within reach in the next few years.

But despite a flood of companies entering the space — bringing tens of billions of dollars in VC and Big Pharma money with them — the field has little hard data to show for itself, at least as far as the public can see. And that begs the question: How much of the “AI revolution” is hype, and how much is grounded in reality?

Conversations with several biotech and pharma executives, as well as investors, suggest there’s a middle ground to be found here, one in which AI can truly change how drugs are discovered and how clinical trials are run. But it’s also one where expectations need to be managed, so companies don’t overpromise and underdeliver.

“Biology is really hard,” longtime AI researcher Daphne Koller told Endpoints News. “There’s just so many interconnected systems in our body where if you poke at one, another one that you completely didn’t anticipate suddenly does something that causes adverse side effects and completely eliminates your therapeutic window.”

Regardless of anyone’s expectations, one thing that hasn’t slowed down is the sheer amount of cash that’s entered the space in the last couple years. Funding rounds well into the nine-figures have become commonplace.

According to numbers compiled by DealForma, there have been 103 venture rounds closed for AI and machine-learning biotechs since 2014, raising a total of $5.1 billion. Almost two-thirds of that, $3.3 billion, came in 2021 alone.

The R&D licensing side, meanwhile, has been even busier, with 243 partnerships signed in that period, totaling more than $1 billion in upfront cash and equity and reaching $37.7 billion in announced deal value. About $10 billion of that came in January of this year.

In a handful of examples, Koller’s own insitro pulled together a $400 million Series C in March 2021, Flagship’s Generate Biomedicines notched its own $370 million round last November and Alex Zhavoronkov’s Insilico Medicine tallied $255 million of its own last June. Companies that have since gone public with massive IPOs, like Exscientia and Recursion, also put together huge private raises before hitting Nasdaq.

Most of these biotechs, though, remain in their preclinical or early clinical stages. There’s little, if any, hard data on whether drugs discovered through AI and machine learning platforms — or those shooting for new targets found through these methods — are actually safe or effective in humans, or whether AI can actually develop better drugs, faster.

And as such programs prepare to read out data in the near-term, there’s a growing wariness that a setback with the first could lead to ramifications for the rest.

The hype surrounding AI is not unique to the life sciences.

Several other industries have undertaken huge efforts to try automating mundane yet crucial aspects of their businesses to cut costs. From retail’s push to self-checkout lines, to automakers’ significant investments in self-driving technologies, even to internet security checks asking you to click which images contain stop signs — the glamour of AI has seemingly seeped into every corner of society.

In biopharma, this push has led to myriad companies launching their own platforms either trying to use neural networks to uncover biological insights previously unseen, or trying to screen a variety of potential candidates faster than humans can.

But what’s largely enabled the industry to even get to this point, Koller tells Endpoints, isn’t the money or a scientific inflection point — even though both have come. Rather, it was the explosion in data collection capabilities and interest in data science that bred the technologies we’re now seeing.

Way back in machine learning’s prehistoric era (also known as the early 1990s), researchers would consider a dataset large if it contained about a few hundred samples, Koller says, and the general goal remains the same as then. Most of the field’s earliest attempts involved teaching machines how to recognize images, with scientists coding in which features they thought would be pertinent.

The tiny datasets, however, made it extremely difficult to train computers to learn on their own.

“What turned out to be the case was that people are just not very good at defining features that actually are the relevant ones,” Koller said. “Like, for example, recognizing objects and natural images. We can all recognize a dog. But it’s really hard to tell the computer how to recognize a dog. And so the computers plateaued at a relatively mediocre level of performance.”

Things began to change around 2012, she noted, when people began to realize how data science could provide a real impact and “Big Data” started to enter the mainstream. Data scientists soon became one of the hottest professions as companies big and small tried to harness all this new information into actionable work.

What that meant, though, was datasets had gotten large enough for researchers to apply machine learning models that could solve complex problems. Whereas the models used to struggle to even identify images, they could now figure out why an image or problem might be important without prompting from humans and make connections on their own.

The implications for biopharma soon became evident and the earliest AI biotechs began popping up. Exscientia, Recursion and BenevolentAI were among that group, with the former launching in 2012 and the latter two in 2013. Insilico followed the year after, but by that point the investor and Big Pharma interest had become clear. Koller herself has attempted to mitigate that problem at insitro, too, which has spent heavily specifically on data collection in order to better train the models.

Vijay Pande

And the money followed, largely from VCs eager to capitalize on the “next big thing.” Vijay Pande, a former Stanford professor who’s now general partner at Andreessen Horowitz heading up biopharma and healthcare investing, said that when he first joined the firm full-time in 2015, anticipation around biology-focused AI applications had already reached a fever pitch.

“It really felt even then that we were at this cusp, where these technologies were showing promise for many years, but when was the right time for them to really start being applied in a corporate sense?” Pande said. “I think that was what I was starting to see then. And so I very much voted with my feet and left academia.”

As the cash flowed in, one thing became eminently clear: The race to develop the first AI-discovered drug was on.

Exscientia and Recursion have both attempted to plant the flag here in similar, yet different, areas. Exscientia, which uses AI to both screen and design new compounds, claims it had the first AI-discovered drug to reach human trials when it launched a Phase I study for its OCD pill in January 2020.

Recursion, meanwhile, uses machine learning to sort through billions of images of diseased and healthy cells in hopes of finding new pathways to combat disease. It made the same assertion as Exscientia six months earlier for two of its programs. But while both came from more traditional discovery methods — one candidate came out of Dean Li’s old lab before he went to Merck and the other was in-licensed from Ohio State — Recursion found new applications for the drugs through its AI platform.

Both were able to translate their progress into fundraising, however. Recursion put together a $121 million Series C that July while Exscientia landed its first major financing with a $60 million Series C in May 2020 after years of partnerships. Both have since gone public in IPOs last year tallying more than $500 million for each biotech at closing.

And they weren’t alone in their promise-making, or fundraising. The California biotech Atomwise was notorious in the mid-2010s for claiming it could do things like shave six years off drug development timelines, sometimes explicitly and sometimes heavily implied.

Alex Zhavoronkov

Zhavoronkov’s Insilico, in particular, has had a knack for generating good press over its preclinical research papers. One such paper, published in September 2019, made the rounds in non-life sciences circles as Zhavoronkov compared the research — in which he claimed Insilico’s AI discovered a drug in under two months — to Alphabet’s Go-playing computer program.

But despite all the hoopla, the only data Exscientia have published come from an open-label, Phase I basket study whose primary endpoint contained numerous qualifiers and compared treatments to a patient’s previous therapy rather than placebo. And Recursion hasn’t published any Phase I data despite saying it’s gearing up for a slate of Phase II studies early this year.

Another biotech, Relay Therapeutics, claims it is the “most mature” AI biotech with three AI-discovered drugs in the clinic. But it’s hard to draw any conclusions from initial data, as they come from a small, six-patient sample reported in October.

Andrew Hopkins

Exscientia CEO Andrew Hopkins said it’s important to keep in mind that many, if not all, AI biopharmas have spent the last few years scaling up their technologies before they could even reach the proof-of-concept stage. His company, he noted, took about nine years to report its first data, from its founding in 2012 to the Phase I readout last October.

Such a milestone itself is a key validator, as it represents the culmination of Exscientia’s progress to this point, he added.

“We have an incredibly powerful example already of how an AI algorithm can improve outcomes in cancer,” Hopkins said. “In fact, that’s the basis of our whole precision medicine approach. We’ve clinically validated that technology now, and that becomes the basis of how we use patient-centric approaches in target discovery, candidate selection and patient selection.”

The hyped-up claims from this first wave have rankled at least a few researchers at some of the newer AI biotechs, including Jen Nwankwo at 1910 Genetics, a Microsoft-backed upstart that emerged from stealth last March. She questions whether the race to be first threatens to leave patients behind. At the end of the day, she says, everyone is still in the business of helping sick people.

Jen Nwankwo

Nwankwo said she views AI and machine learning more as “another tool in the toolbox” to develop good drugs, despite how it could shift the whole R&D paradigm. The interest is definitely real, she said, but whether or not there’s substance behind the hype remains an open question.

“We’re gonna have these vanity metrics like dollars raised, amount of venture capital dollars, size of upfront deals with business development partners and pharma,” she said. “And it’s like, we’re a biotech company, let’s talk about the medicines we’re trying to bring to market. And I think there is not nearly enough focus on that. And to me, I don’t think that’s the right way to think about it.”

Koller, too, feels a lot of the hype stems from the lack of a hard definition on what “AI-discovered” actually means. There isn’t a situation where an algorithm or machine learning model will develop a drug on its own from start to finish. The process is generally an interplay between some AI aspects and more traditional approaches, combined with humans who oversee the whole thing, she said.

The artificial intelligence platforms are definitely assisting the process, Koller said, but “is it assistance of 5%? 50%? 80% of the process? And which parts of the process did it play a role? There’s a benefit to hyperbole that is often hard to kind of tease apart from the reality.”

Exaggerating or not, the money is very real, and doesn’t appear to be stopping any time soon. Recursion signed a gargantuan $12 billion deal with Roche last December to potentially develop up to 40 new drugs, while Exscientia agreed to a $5 billion-plus deal with Sanofi in January.

One of Koller’s biggest worries about overhyping the new AI technologies is that the hyperbole could scare away investors and researchers should the first wave of drugs fail. There have been several instances in the past, she says, of people who made promises so huge that their ultimate flops caused the field to stagnate for up to a decade.

The difference now is the AI technologies are starting to prove they can truly work, delivering real value throughout the R&D process. Computing power is finally starting to catch up to the explosion of data and changes at this point are inevitable, be it at the earliest stages of discovery or in making large-scale pivotal trials more efficient.

What the industry has to do now, Koller says, is “calibrate” people’s expectations to align with what’s possible. She understands the desire to push things quickly — as a longtime AI researcher and venerated titan in the field, Koller wants to see success more than most.

But she feels it’s a desire where excitement has repeatedly outstripped reality, coming at great cost.

“People were dismissive of it, grad students wouldn’t go work in it because it was considered a bit of a laughingstock, and I’ve lived through a couple of those,” she said. “Those were, by and large, because people made hyperbolic promises that didn’t make sense … we run the risk of a similar winter here, and I really think that would be a shame.”

So what can be done to more appropriately manage expectations? There’s no easy answer readily apparent, given how every executive operating in the space is still an entrepreneur, Nwankwo said, including herself.

A potential solution could be to take a page out of Big Pharma’s playbook. Although much of the innovation in the life sciences comes from biotech, the heavyweights are going forward with some of their own AI projects as well. AstraZeneca, GlaxoSmithKline, Novartis and Merck KGaA are among those with some of the more advanced in-house efforts, with AstraZeneca the consensus leader of the group. (AstraZeneca executives were unable to sit down for interviews.)

Kim Branson

Kim Branson, GSK’s global head of AI and machine learning, said when it comes to exploring new technologies, Big Pharma is often much more deliberate than biotech — and that’s been the case in AI too. Pointing toward the deeper pool of resources large companies possess, he said it’s allowed GSK to set reasonable goals while still seeing paradigm shifts.

Novartis’ Iya Khalil, the global head of its AI innovation lab, warned the current machine learning efforts could be likened to the Human Genome Project. The promise of being able to develop dozens of new drugs fizzled once people realized a single genome doesn’t do much to advance those efforts, Khalil said.

AI biotechs still have potential to be different, however.

The key, she said, is taking as many “shots on goal” as possible so that when — not if — a failure occurs, the field can better learn from its mistakes.

“We’re at a place now where we’re having many companies now do this, many biotech companies as well as pharmaceutical companies such as ourselves, take shots at doing this where there’s going to be a variety of approaches,” Khalil said. “And it’s not just about the AI, it’s really about how you make the entire thing work.”

Iya Khalil

Other options include offering services to other companies as a way to remain profitable, something Zhavoronkov has touted sets Insilico apart from his competitors. Though Insilico remains behind on the pipeline front, having only just launched its first trial late last year, Zhavoronkov is trying to outflank other biotechs by selling his company’s software.

“My strategy is, ‘Hey, open up the platform, ensure that people see your tools and can use your tools,’” Zhavoronkov said. That way, companies can also “prioritize programs that are novel, but at the same time, you have high confidence, you have excruciating preclinical evidence that your bet is more likely to succeed than not.”

The one thing most, if not everyone, seems to agree on is this: Biopharma is at the beginning of a period of transformation and change. Whether the magnitude of such change lives up to the hype remains the open question, however. Whereas the 2010s were the decade of data science, the 2020s have the potential to be the decade of artificial intelligence.

“We also don’t want to be a downer and say, ‘Oh, is this all a bunch of hype?’ Because it’s not. I mean, there is so much opportunity in this space,” Koller said.

But there’s a responsibility that comes along with being the first to innovate, Koller adds. There needs to be a keen awareness of what a failure might bring to the field, sort of a pay-it-forward mentality to future scientists, reiterating the failures of the past.

As to what the next few years will concretely bring, predictions range from a steady string of advancements and data readouts to broad adoption of AI technologies across the discovery space. Hopkins went as far as to say that by 2030, nearly all new drugs will be designed by AI or machine learning models.

There are also pushes to integrate patient data into precision medicine approaches and continue to optimize clinical trials, which will likely see huge funnels of cash as well. The Covid-19 pandemic has only seemed to accelerate such advances with new consumer preferences and comfort with telehealth.

Researchers are still a long way from solving everything, Koller emphasized, given how complex these problems are. Anything that moves the needle, however, can prove a difference maker.

“Even if you have a machine learning model that can, in fact, make a big difference in our ability to identify good targets, even if you double the success rate from 5% to 10%, or even to 20, still most things fail,” Koller said. “You really need to kind of have a long view here and understand what we’re trying to do in drug discovery is one of the hardest problems that we face.”

Read More