Have the Promises Made by AI Experts Come to Fruition?

An update on Eureka’s AI in Drug Discovery series. 

Four years ago, Eureka ran a series on the use of artificial intelligence (AI) and machine learning (ML) in drug discovery, covering the legal, ethical, and practical applications of the emerging technology. In computer science years, four might as well be forty, so we were naturally curious to find out what change – and what hasn’t—in this fast-moving industry.

Practical Applications of AI in Drug Discovery

In 2020, just one year after our series ran, the first AI created drug candidate entered human trials. The drug, which was designed to treat obsessive compulsive disorder (OCD), was brought to this stage in only 12 months, though trials are currently still ongoing. Since then, the same company (Exscientia) reports three other AI-designed drug candidates in Phase I trials. 

One of the biggest developments in machine learning has been AlphaFold, developed by the AI research laboratory Google DeepMind to bring the power of AI to the protein folding problem. Predicting how proteins fold is a key step in small molecule drug discovery, and another potential benefit of bringing AI into the discovery lab.

“I think AI has actually exceeded expectations in some senses, particularly in the area of protein structure prediction (AlphaFold etc.),” said David Clark, Senior Research Leader for Charles River’s Small Molecule Drug Discovery group in Harlow, UK. “I don’t think anyone saw that coming, and AF has really been the poster child for AI in the life sciences.”

Has AI Lived Up to the Hype?

While AI applications in drug discovery are steadily growing, is it living up to its reputation and promises? 

“The answer to this is perhaps dependent upon the expectations that were held four years ago,” said Grant Wishart, Senior Director and colleague of Clark’s. “Whilst the unrealistic view that AI/ML would provide a magic solution to solve all our drug discovery challenges has clearly not come to fruition, what we have seen is a growing presence of AI/ML in drug discovery and a greater understanding of where it can impact.”

Wishart refers to the unrealistic expectation held by some that AI could replace human researchers entirely, and spontaneously create perfect drug candidates that can go straight to shelves. Besides being inaccurate, this expectation can hinder enthusiasm for AI by overshadowing the very real contributions it can make to the drug development process. 

Wishart has seen for himself the value of AI in small molecule drug discovery including target identification, predictive models of on and off target binding, and (as Clark mentioned) predictive protein structures. What AI may lack in true creative ability, it more than makes up for in its ability to consume and synthesize vast quantities of data into a form that is useful for scientists, like ranking and scoring vast numbers of compounds for binding to a protein target of interest. Ranking candidates like this saves precious human work hours by allowing researchers to focus on the most promising candidates.

“Drug discovery is data driven where we perform experiments, analyse the results, learn, then plan the next experiments,” said Wishart. “How we achieve this is constantly evolving, none more so than in recent years where the utility of AI/ML has been explored and the potential impact is starting to being realised.”

AI hasn’t just affected drug discovery. Chemists are also benefitting from AI’s ability to synthesize piles of data into byte-sized nuggets of information. According to theoretical chemist Dr. John Mitchell from the University of St. Andrews, the excitement for and investment in AI has not abated in the past four years.

“Something that took me by surprise is how pervasive ML is now in chemistry,” he said. “We’re seeing ML routinely used to process and analyse experimental data from techniques such as spectroscopy and NMR. For example, in my own University, about a third of the Chemistry research groups, across all disciplines of the subject, are now using ML or AI in some form.”

Ethical and Legal Implications of AI

This is not to say, however, that every question has been answered. Mitchell especially stated that it might be too early to tell the long-term impacts of AI/ML on drug discovery, since any research in this industry takes time. Furthermore, each new advance in AI in any industry brings a new onslaught of ethical and legal questions to answer.

“AI art (for example) went very quickly from being a fun toy to being a major intellectual property (IP) dispute,” said Mitchell. “As a relatively mundane example, within the academic community, there have already been cases of AI images allegedly being used in scientific fraud and data fabrication. In a field like pharma, where any question marks over the ownership of a project’s IP could be existential risks to the company, there will be some well-remunerated IP lawyers.”

Beyond the sticky legal implications of who owns the fruits of AI’s creative genius when it is trained with multiple data sources, there are also issues ranging from privacy and surveillance to data security. On the ethical angle, there are many debates over how to eliminate explicit and implicit human bias in machine learning – when the computer is being fed information by people, it is inevitable that it could also be getting a hefty dose of the human trainers’ biases.

“Humanity has survived the transitions wrought by the printing press, productions line, computer, and internet,” said Mitchell. “Nonetheless, it’s very reasonable to question whether transformational technologies are indeed beneficial, and anyone who’s read 1984 may have concerns along the lines of ‘cell phone-as-telescreen’.”

“It’s impossible to dispute the rapidly growing footprint of AI – it seems to be cropping up almost everywhere,” said Clark. “It will be interesting to see how everything shakes down over the coming years and how quickly we reach a “new normal” where AI is an accepted tool in the box, intelligent assistant or something more.”