Artificial intelligence (AI) has long been hailed as a game-changer in drug discovery, promising to accelerate the development of new treatments. More than 150 drugs are in development using AI, but only one has been approved, and even then, AI had a minimal role. Challenges such as regulatory uncertainty and the need for vast, high-quality data are slowing progress. However, experts think that AI still has untapped potential, with future breakthroughs possibly just around the corner as collaboration and clearer regulations advance.
AI has been a hot topic in drug discovery for more than a decade, promising to accelerate the discovery and approval of new therapeutic agents. But this potential has yet to bear fruit, according to a recent analysis published in JAMA. Although AI has contributed to more than 150 drugs under development, just one drug has been approved for humans – and in that case, the AI was only used to analyse the statistics of the clinical trials.
“I do not think that the promise of AI in drug development has peaked,” says Louise C. Druedahl, who co-authored the study as a Postdoctoral Fellow at the Centre for Advanced Studies in Bioscience Innovation Law, University of Copenhagen, Denmark. “I believe that we will see much more in the years to come.”
“But an uncertain regulatory landscape for AI in drug development could hold back further progress,” says co-author Timo Minssen, Professor at the University of Copenhagen who studies the law and ethics of emerging health technologies.
AI in early drug development
By scanning a database of press releases from drug companies, Druedahl and co-authors identified 164 drugs being developed whose makers report using AI at some stage. About one third of these drugs aim to treat people with cancer and another third target people with neurological conditions, Druedahl says.
AI was most frequently used to assist in discovering drug molecules, the researchers found, accounting for 125 of the 164 drugs identified (76%). Screening potential candidates for therapeutic applications is very much a numbers game, and AI can help to process the large quantity of data to evaluate molecules or make benchwork unnecessary altogether.
“The approaches seem to vary, since some examples include platform screening of drugs, using AI to analyse molecular images of how various drugs affect a cell, and in others deep generative modelling was applied to design virtual novel molecules,” Druedahl explains.
The second-leading use was discovering drug targets – drug manufacturers reported that machine learning has helped to make connections between “genetic, chemical and clinical data” that might have otherwise gone unnoticed. This represented 37 of the 164 drugs (22%).
The single AI-assisted drug that has reached the market is remestemcel-l, a stem cell therapy. However, remestemcel-l is not much of a model for the power of AI to accelerate drug development – the AI was only used to analyse data after the clinical trial was complete. “A Bayesian method was used to estimate the likelihood of obtaining significant results on the primary end point at study completion,” the researchers wrote.
One limitation of the study is that it only captured cases in which drugmakers disclosed using AI in press materials. “We would not capture in the study if manufacturers use AI without communicating about this publicly,” Druedahl says.
What is holding back AI in drug development?
Several hurdles need to be overcome to realise the potential of AI in drug development, Minssen says.
Training AI for drug development will require tremendous stores of high-quality data – more than any one drug manufacturer currently has access to, he says. “Companies will probably have to collaborate and share data” to reach that critical threshold, and “new forms of coopetition, or cooperation between competitors, must be considered,” Minssen adds.
This would require adjusting existing laws and company policies governing trade secrets, intellectual property rights and data protection laws that cover medical data.
But all data are not created equal. “In drug research and development, data often come from diverse sources such as clinical trials, preclinical studies and real-world evidence, which can be incomplete, inconsistent or noisy,” Minssen says. “Standardising data collection and curation processes across the industry is essential along with developing robust data integration frameworks to harmonise and clean data from various sources to increase interoperability and usability.”
Regulatory uncertainty and ethical considerations
Another stumbling block is that the legal standards for AI in drug development have not yet solidified. Both the United States Federal Drug Administration (FDA) and European Medical Agency (EMA) require “rigorous validation” of AI tools, which can be “time-consuming and resource-intensive,” Minssen says.
“More clarity is needed for companies not only on what level of proof is expected but also how this can be demonstrated and achieved,” he adds.
It does not help that many AI models, especially deep learning models, are black boxes with opaque decision-making processes. This can understandably make regulators reluctant to rely on their conclusions.
“Demonstrating the reliability and reproducibility of AI models is crucial for regulatory approval,” he says. “Publishing detailed protocols and results to facilitate reproducibility can also help to build trust and credibility with regulatory authorities.”
The FDA and EMA are currently re-evaluating their guidelines and procedures on using AI in drug development. Minssen says that he hopes that collaboration with industry leaders early in development could help to “streamline the approval process”.
The ethics of using AI in drug development will also require careful tending, Minssen says. “AI models can inherit biases from training data, leading to unfair or biased outcomes. Using diverse and representative data sets for training AI models and implementing bias detection and mitigation techniques are important steps,” he explains.
“Ensuring that AI is used ethically in drug research and development, particularly in decisions that affect patient health and safety, requires establishing ethical guidelines and frameworks,” Minssen adds.
“Creating oversight committees to review AI applications and their impact and promoting transparency in AI decision-making processes can help to ensure ethical use,” he says.
What is coming down the pipeline?
“AI has the potential to revolutionise drug research and development,” making the process “more efficient, accurate and cost-effective,” Minssen says.
Since 2020, the use of AI in drug development has shifted from primarily “repurposing” existing molecules to treat people with various conditions to “discovering” new drugs, according to an analysis published in Drug Discovery Today.
Forty-six “AI-discovered” drugs – in which AI identified a new treatment target or drug molecule – reached Phase 2 or 3 clinical trials in 2023, the analysis found. This means that the first truly AI-discovered drug could be on the horizon – and one day soon, in a pharmacy near you.
