*Guest blog by Professor Alan M. Palmer, Chief Executive Officer, Elixa MediScience.
If you would like to submit a guest blog please email [email protected].
The industry promised AI would transform medicine. Instead, it delivered a speed boost nobody needed and a clinical failure rate nobody fixed.
The pharmaceutical industry has spent the better part of a decade being told that artificial intelligence would transform drug discovery. The pitch was compelling: AI would slash development timelines, cut costs, and defeat diseases that had resisted human ingenuity for generations. Billions of dollars flooded into a new generation of AI-native biotech companies. Share prices soared. Breathless headlines followed.
Then reality arrived.
The story of AI in drug discovery is not one of outright failure—the technology has delivered genuine, measurable value. But it is a story of a profound and increasingly expensive mismatch between what AI was promised to do and what it has actually managed to achieve. Understanding that gap is essential for anyone trying to think clearly about where this technology goes next.
Where AI Actually Works
To be fair to the technology, AI tools do deliver real value in the early, preclinical stages of drug development. This is the phase where researchers scan vast chemical spaces to find candidate molecules—compounds with the right shape, charge, and behaviour to interact with a biological target in a therapeutically useful way. AI can generate enormous numbers of potential compounds rapidly, helping scientists narrow down their search far faster than traditional methods allow.
The results are real. Insilico Medicine, one of the more credible players in the space, took an AI-discovered drug candidate for idiopathic pulmonary fibrosis—a severe and poorly treated lung disease—from target identification to Phase II clinical trials in under 30 months. The same journey through conventional means typically takes many years. That is a genuine achievement, and the industry is right to acknowledge it.
But there is a crucial distinction buried in that success story. Getting a candidate to the starting line of clinical trials is not the same as getting it across the finish line. And it is in the gap between those two points that the AI drug discovery paradox begins to reveal itself.
The 85–90% Problem
Clinical trials fail. They fail at staggering rates—somewhere between 85 and 90% of all drug candidates that enter human trials never make it to approval. Compounds that look promising in the lab turn out to be toxic, ineffective, or simply irrelevant once tested in real patients. This attrition rate has been one of the pharmaceutical industry’s most persistent and costly problems for decades.
The central promise of AI was, implicitly or explicitly, that it would improve these odds. Smarter candidate selection, better target validation, more sophisticated modelling of human biology—all of these were supposed to translate into drugs that were more likely to work in people.
Early data suggests this hasn’t happened. A drug discovered with AI in record time appears, on current evidence, to fail in the clinic just as readily as one found through traditional means. Speed to the starting line has improved. The race itself remains as brutal as ever.
The Broken Value Proposition
The consequences of this gap between promise and reality have been severe, and not only for the companies involved.
The hype around AI drug discovery dramatically inflated expectations and valuations throughout the sector. A 2023 Financial Times report noted that while billions had been poured into the space, no AI-discovered drug had yet reached late-stage trials—a clear failure to deliver on the promises that had justified those investment levels. What followed was predictable, if painful: several pioneering companies collapsed or were forced into fire sales at a fraction of their peak valuations.
After going public via a SPAC merger on Euronext Amsterdam in April 2022, BenevolentAI was valued at approximately €1.5 billion (£1.3 billion). Now its share price has fallen by more than 99% before being delisted. Exscientia, another high-profile name, was acquired for a fraction of its peak valuation after its AI-designed cancer drug candidate, EXS-21546, failed in mid-stage trials. The human cost of these failures—in capital destroyed, in promising scientific programmes abandoned, and in patients who might have benefited from treatments that never materialised—is substantial.
One industry chief executive, speaking with unusual candour, summarised the prevailing sentiment: “AI has really let us all down in the last decade when it comes to drug discovery. We’ve just seen failure after failure.” That is a damning verdict from inside the tent.
The Biological Data Problem
To understand why AI has struggled to deliver on its clinical promises, it helps to understand the fundamental mismatch between how AI systems work and the nature of biological systems.
AI models excel at finding patterns in clean, structured, high-quality data. Biology, unfortunately, is none of those things. It is messy, multi-dimensional, and heavily context-dependent. Biological datasets are frequently noisy, riddled with gaps, and radically incomplete relative to the complexity of what they are trying to describe. A molecule that behaves in a particular way in a cell culture behaves differently in a tissue, differently again in an organ, and differently still in a living, breathing human being with their own genetics, microbiome, diet, and disease history.
As the respected pharmacologist and drug discovery blogger Derek Lowe has noted, AI needs “much, much more knowledge of human biology of health and disease” than we currently possess. The models are only as good as the data they are trained on, and the data we have is nowhere near adequate to capture the full complexity of human physiology.
This limitation extends even to tools widely celebrated as AI breakthroughs. AlphaFold, DeepMind’s protein structure prediction system, was rightly celebrated for solving one of biology’s great computational challenges. But its relevance to drug discovery has been tempered by significant limitations that are sometimes glossed over in the popular coverage. The system can produce deceptively confident predictions that turn out to be inaccurate when it comes to modelling the complex, dynamic interactions between a potential drug molecule and its target protein. More troublingly, some research suggests that AlphaFold may work partly by matching inputs to a stored database of known protein structures—effectively pattern-matching from memorised templates—rather than by understanding the underlying physical and chemical principles from first principles. For discovering truly novel drugs that interact with previously unknown targets, this is a critical flaw.
The Functional Discovery Gap
Even setting aside the data problem, there is a broader challenge that AI has not yet begun to address. Identifying a promising molecule is one step in a vastly longer and more complex journey. Between a candidate compound and an approved therapy lies a gauntlet of challenges that no AI system currently knows how to navigate.
These include understanding and predicting complex immunological responses, designing clinical trials that are ethically sound and statistically robust, assessing real-world outcomes across diverse patient populations, manufacturing compounds at scale, and navigating the regulatory frameworks that exist to protect patients from harm. AI can suggest a molecule. It cannot yet shepherd that molecule through the high-dimensional, deeply human process of turning it into a safe and effective medicine.
This is not a criticism of AI’s potential. It is a realistic account of where the technology is today.
The Uncomfortable Conclusion
AI has made early-stage drug discovery faster. That is real, and it matters. The Insilico Medicine example is not an anomaly—there are genuine efficiency gains being captured across the industry in lead generation, compound optimisation, and target identification. These gains will compound over time as the technology matures.
But faster is not the same as better. The fundamental bottleneck in pharmaceutical development has never been the speed of preclinical candidate generation. It has always been clinical attrition—the brutal, expensive, and often heartbreaking process of discovering that compounds that look promising in the lab fail in people. Until AI demonstrably improves clinical success rates, the efficiency gains at the front end of the pipeline are, at best, a partial solution.
The key question, as one industry commentator has put it, is not whether AI can accelerate preclinical timelines—it clearly can—but whether it can improve the odds of success in the clinic. Until we have robust data to answer that question, the pharmaceutical industry’s increasingly cautious approach to AI investment looks not like timidity or Luddism, but like hard-won wisdom.
The race is still very much on. AI may yet prove transformative in ways the current evidence doesn’t yet capture. But the hype cycle has done its damage, and the next phase of AI drug discovery will need to be built on something more durable than optimism: rigorous data, honest assessment, and a clearer-eyed understanding of the problems that actually need solving.