In the mid-2010s, a spate of start-ups hoping to transform the laborious process of finding new drugs launched with big promises. Artificial intelligence would dramatically reduce the time it took to discover new medicines and cut the average of $2bn it takes to develop a drug.
The emerging businesses attracted the attention of Big Pharma companies such as Bristol Myers Squibb and Sanofi, which signed deals worth billions of dollars pending the drugs' eventual approval. Press releases boasted of "breakthrough productivity gains" and "groundbreaking research collaborations".
But now, sceptics are asking: where are the drugs? It has been longer than the average 10 years that it takes to discover and develop a medicine, yet there are few AI-discovered candidates in late-stage clinical trials, and not one has been approved. Despite pledging to cut the industry's high failure rate, many of the companies' initial studies flopped.
Some start-ups have struggled financially, launching in a period where investors have pulled money from the biotech sector in general. BenevolentAI, a British company that attracted lots of early excitement, saw its shares fall more than 99 per cent before it delisted in March, merging with a Japanese company. US-based Recursion snapped up rival Exscientia cheaply last year, paying $688mn -- only $180mn more than the cash on its balance sheet and far less than the $2.9bn valuation it went public at three years before.
Alex Zhavoronkov, chief executive of AI for drug discovery company Insilico, says companies have been under pressure to prove their big claims about transforming drug discovery, by showing they had actual drugs. "In order to claim that you have a golden goose, you need to make sure that you have laid a few golden eggs. And if you don't have golden eggs, your golden goose is depreciating very, very quickly," he says.
Daphne Koller, chief executive of another, similarly named, AI for drug discovery start-up, Insitro, says fundamentally we are trying to fix something we do not understand because of the complexity of human biology. "I used to say we were the industry with the highest failure rates of anything but space exploration. And then space exploration started to work," she says.
The idea of applying AI to drug discovery has been so attractive because investors see the pharmaceutical sector as a key area where slow and expensive processes are ripe for disruption. Venture capitalists poured money into companies trying to use AI to discover new drugs, seeing the industry as a promising field, especially as ageing populations increase medical costs around the world. Funding for AI drug discovery companies increased from $30mn in 2013 to a peak of $1.8bn in 2021, according to data from PitchBook.
The meteoric rise of generative AI since the launch of ChatGPT in late 2022 has kick-started a new boom in using powerful AI technologies to design drugs.
AI tools are already showing potential in other scientific fields, such as extreme weather prediction. Investors are being lured back in by hope that a fresh crop of companies, technological breakthroughs and new ways to collect and understand biological data could finally lay a golden egg.
The bet behind this new generation of companies is that AI can still revolutionise the drug discovery process, but that the initial tools being used were not powerful enough. As a result, supporters say, it is too soon to rush to judgment.
Yet so far, the problem has proved beyond the algorithms. We know surprisingly little about our own biology. There are many mysteries in how our cells interact and challenges in measuring our body's most crucial processes, starving models of the data they need to make more rapid progress.
Darren Green, a veteran chemist who spent more than 30 years at GSK, says drug discovery is "probably the hardest thing mankind tries to do". "We get these great new tools, which is fantastic. And then you just find another problem," he says.
AI was not the first technology that the pharma industry anticipated would unlock the secrets of life. Similar hopes were pinned on structural biology in the 1950s and 1960s, as scientists used tools like X-ray crystallography to understand proteins in 3D; computational chemistry, which simulated experiments in the 1980s; and the discovery of the human genome in the early 2000s.
In the drug discovery process, scientists typically identify a target in the body, such as a mutation in a tumour or a receptor for a particular hormone, and then search for a molecule that can hook on to it and change its behaviour to address a symptom or a disease. Researchers have to design the compound so that it hits the target, and does not cause havoc in other parts of the body. Drugs that look good on paper still fail about 90 per cent of the time in clinical trials.
The appeal of AI is that it can speed through databases of molecules, matching compounds to targets. However, that is just one element of drug discovery, and many say one of the easiest. Peter Coveney, a professor and director of the centre for computational science in UCL's chemistry department, says that toxicity -- side effects from a drug -- can be particularly hard to predict.
"It is a false picture to imagine any whizz technique with computers is just going to solve all the problems," he says.
That's why regulators require new drugs to undergo multiple stages of testing -- first in animals and then in humans -- a process that accounts for the longest part of the typical decade-long journey to approval.
Koller says that many of the successful applications of AI today are "[computer] bits meet bits", such as ChatGPT learning from language on the internet, or AlphaGo playing the Chinese strategy game Go.
But drug discovery is a trickier challenge, because it operates where "bits meet atoms". She compares it to self-driving cars, where adoption cannot be accelerated by faster chips. "We're finally getting to a point where you have self-driving cars . . . it took a long time, because these things take a long time," she says.
In 2013, at about the same time as BenevolentAI and Exscientia were founded, Chris Gibson co-founded Recursion. The company has four potential drugs in early-stage trials in oncology and rare diseases, but nothing yet in the final and most important stage that it would need to get an approval.
He is still a true believer in the power of AI to change drug discovery, but is disappointed the field has not come further. "The ones that were around when we started have either gone, or we bought them," he says.
Kenneth Mulvany, who founded BenevolentAI in 2013 and recently came back to lead the company, says the technology in the early years was moving so quickly that it looks very different from the way that AI is used today for drug discovery. "Each year, when we would build something, we would have to supersede that with something else," he says.
The company would spend an enormous amount of time juggling thousands of different databases and using algorithms to analyse and explain their content. "Getting all of that information was actually quite a difficult task," he says.
The public datasets that were available at the time were comparably small and even large pharmaceutical companies, which pride themselves on having decades' worth of data from testing and trials, often had it scattered in separate spreadsheets with too much noise to be useful for AI algorithms.
Earlier generations of companies would use machine learning models to develop a specific tool for each problem, says Miles Congreve, the chief scientific officer of Isomorphic Labs, the drug design spin-off of Google's AI unit, DeepMind. "It was quite hard for those companies to really impress people [because] you've solved this problem with your AI platform, but it's not generalisable to the next problem," Congreve says.
The early generation of AI start-ups also had knotty strategic choices. Ideally, they would invest in many potential drugs at once, to try to prove they had a greater success rate than traditional methods. But few could access that money and so the choice of the first target was crucial.
One industry executive who wants to remain anonymous says the "ultimate proof" of AI's transformative power would be if these companies could show success where pharma consistently failed, such as treating Alzheimer's, or tackling brain tumours. But, he adds, instead many chose easy targets, under pressure to show success quickly and appeal to the potentially risk-averse buyers in Big Pharma. Sometimes known as "me too" or "me better" drugs, they selected well-known biological targets and focused their AI on finding new compounds.
The potential drugs that companies found often were not that much better than those already on the market, he adds.
Sanjiv Patel, chief executive of Relay Therapeutics, which is trying to develop drugs for difficult-to-tackle diseases, says many rivals went out of business after failing to get a drug to a trial. "The companies that produced 'me toos' realised very quickly that there's no value in that," he says. "Therefore that then leads them back to the beginning again, and they run out of money."
Some of the technology investors attracted to the field were not ready to take the binary risk required by backing a drug in an expensive trial that could fail. Other companies that were dominated by biotech investors were encouraged to invest only in their potential drugs.
Mulvany says BenevolentAI ended up being led by people with experience in pharma who did not prioritise the technology investment. "They were like professional football athletes. They know how to work as a team. They have endurance. They know how to handle the ball," he says. "And you just pick those people up and you pop them on a basketball court and say: 'You're a team, that's a ball.' But there's different objectives and roles to play."
While the older generation of companies have struggled to make good on their promises, many in the industry say the clock restarted with two significant moments.
One was the release of the Nobel Prize-winning protein-folding prediction engine AlphaFold2 in 2021 by Google DeepMind and Isomorphic Labs. The second was the explosion of generative AI starting in 2022. It will still be many years until drugs designed after these two advances are ready for approval.
Gibson from Recursion says the pharma sector is lagging behind others in adopting AI. But he is starting to see signs of a change that could become a "really important" shift for the industry. "When it finally happens, it's going to be one of those slowly and then all at once moments where this is the way everybody has to really start doing discovery and development," he says.
One significant advance is that there is more data available than before. One of the key things that made AlphaFold possible was a vast and well-labelled database of proteins that already existed.
Now many of the surviving start-ups and the new generation of companies are doubling down on data creation. Insitro has created "cell factories", where it uses lab machines to change and edit cells and record everything that happens inside. Recursion uses computer vision on images of human cells and believes it has an internal dataset a thousand times bigger than the largest public database. Lila Sciences, founded in 2023, is using what it calls "AI science factories" -- autonomous labs that it hopes will produce new scientific knowledge -- to fuel its discovery platform.
Companies and academics are also collaborating to collect better data, such as the UK's OpenBind consortium launched this year to use experimental technology to create the world's largest collection of data on how drugs interact with proteins in the body.
The arrival of AlphaFold 2, which lets companies predict how proteins fold and helps scientists improve their understanding of targets for drugs, was a "watershed moment", says Max Jaderberg, chief AI officer at Isomorphic. AlphaFold2 showed you could have algorithms that can generalise across biology. The company has made it freely available for research.
The generative AI boom has also benefited drug design, where scientists try to improve on an initial discovery to ensure that a compound is as safe and effective as possible. Start-ups are now leveraging generative technologies, such as image generation, in their drug design.
Isomorphic Labs uses AI tools that can predict structure beyond just proteins. Powerful generative models, like those used for image creation, design very specific molecules -- a formerly slow and laborious task. It also means that instead of having to create a new model for each specific new problem, the researchers can have one that can be applied to many different targets.
"You can get a model that is trained on some portion of data, but then you can apply it to everything else out there in the universe that can possibly pop up, and it gives you meaningful answers," Jaderberg says.
He thinks this is just the start. To create a true drug discovery engine, he believes we need half a dozen AlphaFold-level breakthroughs. These could include understanding not just the structure of proteins but how strongly molecules bind with targets, how drugs interact with different parts of the body, or predicting the dosage of drugs for patients.
While Isomorphic is keeping much of its work under wraps, Jaderberg describes an ambition far from the "me too" drugs of the previous generation. Swiss drugmaker Novartis, which is partnering with Isomorphic, gave the company "very hard" targets to work on, according to Congreve. Internally, the company has focused its work on cancer and immunology.
Founded in 2021, Isomorphic still does not have any potential drugs in clinical trials. But it does not have to worry about its funders abandoning it after one failed project given that the company is part of Alphabet, the owner of Google. This year it raised $600mn in a round led by US venture capital firm Thrive Capital.
Congreve, who previously worked for GSK and in biotech, says he was attracted to the ambition of Isomorphic and its healthy coffers. "The funding is not all directed at projects. It is directed at making a generalisable drug design engine, which requires data, that requires a lot of compute [power]," he says. "Most smaller AI companies wouldn't have the compute power to build the sorts of models that we do."
While start-ups pioneered the field of AI for drug discovery, it may be large technology companies with money and compute power that end up creating the first successful drug.
As Relay's Patel puts it: "Only those companies could sit on multiple decades' worth of data generation with no value creation, knowing that, in the end, something good can come out."
