When artificial intelligence (AI) entered drug discovery, expectations soared, and bold claims followed: AI would slash development timelines, tame astronomical failure rates and finally unlock cures for diseases such as cancer, dementia, and even ageing. A decade on, the reality looks more complicated. There is no shortage of technical success stories. AI has demonstrably accelerated early-stage drug discovery, compressing timelines that once spanned years into months. Companies such as Insilico Medicine and Recursion have shown that machine-learning can rapidly identify and advance candidate molecules through preclinical development.
Yet for all this progress, a sobering fact remains: not a single AI-discovered drug has completed clinical trials and reached patients. That disconnect, between faster discovery and unchanged clinical outcomes, sits at the heart of a growing reassessment of AI’s role in drug development. In a Nature article, Rachel DeVay Jacobson, co-founder and chief scientific officer of Powerhouse Biology, argues that while AI has delivered efficiency gains, it has failed to meaningfully improve clinical efficacy. The bottleneck, she and others suggest, was never molecule generation alone. Her article, entitled “The AI drug revolution needs a revolution”, outlined how AI has failed to deliver on several promises in drug design. “When the AI revolution marched into the drug development sector and flipped biotechs into techbios, it came with the hope that therapeutic programme failure rates would fall and the holy grails of cancer, dementia and even ageing would finally be found,” she wrote.
“We have seen several success stories where AI substantially accelerated preclinical development, including Insilico’s impressive 30-month sprint and their reported average timeline of 12-18 months to advance 22 benchmark programmes from discovery to Investigational New Drug (IND)-
enabling studies.
“Recursion also reported a speedy advancement of REC-1245 from discovery to IND-enabling studies in just 18 months,” she noted. “While cutting time (and costs) is commendable, we have yet to see these accelerated timelines translate into revolutionary improvements in clinical efficacy.”
Professor Peter Coveney, Director of the Centre for Computational Science (CCS) at UCL (read our interview with him on page 12), believes AI is often treated as a magic bullet – find the right molecule and everything else will fall into place. However, drug development is an arduous process characterised by clinical uncertainty, biological complexity and regulatory hurdles that AI addresses only marginally. The result is a field forced to confront a difficult question: has AI been solving the wrong problems?
Coveney recently said on a Science Friday broadcast that no AI drugs have yet reached the market. While several success stories have been reported, such as Insilico Medicine and Exscientia, as Coveney acknowledges that no AI drugs have yet made it through clinical trials. He said AI may have accelerated the selection of a new candidate drug or preclinical development, but discovery is not the primary obstacle preventing new drugs from reaching the market.
The clinical timeline and failure rates are largely unaffected by the rise of AI in drug discovery. “There are drugs going through the approval process. But the whole point of this story is that drug discovery is an extremely long and torturous process, and things may have been discovered by AI, and that’s a debatable issue in its own right, but it’s not as if you discover something and then it immediately gets approved. New drugs have to go through all these additional hoops, which are things that AI, typically, has absolutely nothing to do with,” said Coveney.
The field has experienced notable challenges that provide essential insights into the current limitations of AI in drug discovery. While AI can accelerate the discovery and optimisation phases, clinical success depends on many factors beyond initial compound design. The discontinuation of DSP-1181 after Phase I, despite demonstrating a favourable safety profile, illustrates that accelerated discovery timelines do not guarantee clinical progression. The reasons for discontinuation often involve complex factors, including efficacy endpoints, competitive landscape and strategic business decisions, rather than issues with the AI-designed molecule itself.
In a paper published in the medicinal chemistry section of MDPI, entitled: “Artificial Intelligence in Small-Molecule Drug Discovery: A Critical Review of Methods, Applications and Real-World Outcomes”, lead author Professor Sarfaraz Niazi of the University of Illinois backs up the points made by Coveney but notes that it is too early to tell if AI has sped up clinical trials.
“A comprehensive analysis of AI- discovered and AI-assisted compounds reveals that, while these molecules are entering clinical trials at an increasing rate, their progression rates through clinical development remain like those of traditionally-discovered compounds,” Niazi writes.
“This finding suggests that AI’s primary benefit may lie in accelerating the preclinical discovery phase rather than fundamentally changing the probability of clinical success. The limited number of AI-designed compounds that have completed clinical development makes it premature to draw definitive conclusions about their overall success rates compared with traditional methods.”
Another point highlighted by DeVay Jacobson is that efficacy has not improved significantly. This raises the question: should pharmaceutical companies use AI to deepen their understanding of biological mechanisms, rather than simply selecting new
candidate molecules? Part of this is due to the relatively small number of AI-first molecules that have yet to reach the clinic, and among those, the majority act on previously established targets. While safety and tolerability tend to improve relative to traditionally developed molecules, their mechanisms of action are generally comparable to those of previously approved drugs.
“There is not a sufficient number of cases to power statistical analyses yet, but if these first few AI drugs serve as proverbial canaries in the coal mine, our current AI approaches may not do much to improve clinical efficacy,” stated DeVay Jacobson. “Thus, while there have been AI-led improvements in programme speed and safety, AI efforts, so far, have
not resulted in more effective drugs.”
Despite initial progress and success in discovering new molecules, Niazi noted that several challenges continue to limit the effectiveness of AI in small-molecule drug discovery. He said: “Data quality represents perhaps the most critical limitation, as AI models are only as good as the data used to train them. Pharmaceutical datasets often suffer from systematic biases, missing data and inconsistent experimental protocols that can lead to poor model generalisation.
“Model interpretability remains a significant concern, particularly for regulatory applications where understanding the basis of AI predictions is crucial for safety assessment. While explainable AI methods have made progress in providing post-hoc explanations for model predictions, truly interpretable models that provide mechanistic insights remain elusive. The “black box” nature of many deep learning models creates challenges for medicinal chemists who need to understand and trust AI recommendations. ”
Quantifying uncertainty
“The common misconception is that AI can act as a magic bullet: find a molecule, and everything else will follow. AI is good at searching for new molecules, but that is only the ‘base camp’ of drug discovery,” noted Coveney. “Many other stages have little to do with AI. In some cases, AI may reduce the overall timeline from 10 years to nine years. Valuable, yes, but far from the headline claims that AI can deliver new drugs in as little as two years.
“The problem is compounded because AI is claimed to apply to virtually every domain, while scientific applications demand reproducibility and reliability. Take AlphaFold: despite its Nobel Prize, it predicts protein structures only under specific circumstances for sequences it has seen. For new sequences, the model’s predictions may not be reliable. The challenge is that we need to quantify the uncertainty in AI predictions rather than rely on them. This is an open problem in science and medicine, and addressing it would significantly improve AI’s credibility.”
Many scientists fail to understand the uncertainty in molecular dynamics simulations and the discipline of uncertainty quantification has emerged over the past decade to address this. This provides error bars or confidence levels for in silico predictions, which is crucial in applications ranging from weather forecasting to healthcare.”
“The key point: AI models can fit almost any data, but we cannot currently assess the reliability of predictions on new inputs,” added Coveney. “Even if a small subset of parameters dominates uncertainty, those parameters are purely fitting parameters; they do not encode physical or chemical information. We can generate outputs we hope are correct, but cannot verify their reliability.
“In theory, uncertainty quantification could make AI more reliable, but doing so would require exascale computing. Short-term, AI has a limited role in drug discovery: generating hypotheses or suggesting molecules, but not replacing experimental validation or regulatory processes.”
Antibiotic design
However, it’s not all doom and gloom. As Coveney notes, AI can be a useful tool in the computational chemist's arsenal; it should be used appropriately, with realistic expectations, as an exploratory tool in the design of new drugs.
MIT Professor James Collins is one of the founders of synthetic biology and is a leading researcher in systems biology, the interdisciplinary approach that uses mathematical analysis and modelling of complex systems to better understand biological systems. His research has led to the development of new classes of diagnostics and therapeutics, including for the detection and treatment of pathogens such as Ebola, Zika, SARS-CoV-2, and antibiotic-resistant bacteria.
Collins, the Termeer Professor of Medical Engineering and Science and professor of biological engineering at MIT, is a core faculty member of the Institute for Medical Engineering and Science (IMES), the director of the MIT Abdul Latif Jameel Clinic for Machine Learning in Health, as well as an institute member of the Broad Institute of MIT and Harvard, and core founding faculty at the Wyss Institute for Biologically Inspired Engineering, Harvard.
His recent research has explored the use of generative AI to design new antibiotics. In a recent Q&A with MIT News, Collins explained that the team used genetic algorithms and variational autoencoders to generate millions of candidate molecules. “After computational filtering, retrosynthetic modelling and medicinal chemistry review, we synthesised 24 compounds and tested them experimentally,”
Collins stated. “Seven showed selective antibacterial activity. One lead, NG1, was highly narrow-spectrum, eradicating multi-drug-resistant Neisseria gonorrhoeae, including strains resistant to first-line therapies, while sparing commensal species. Another, DN1, targeted methicillin-resistant Staphylococcus aureus (MRSA) and cleared infections in mice through broad membrane disruption. Both were non-toxic and showed low rates of resistance.”
He said that the research aims to explore the use of deep learning to design strong antibiotic candidates for clinical development by selecting compounds with “drug-like” properties. “By integrating AI with high-throughput biological testing, we aim to accelerate the discovery and design of antibiotics that are novel, safe and effective, ready for real-world therapeutic use,” said Collins.
