AI and Drug Discovery: How Artificial Intelligence Is Accelerating Medical Research in the UK

Developing a new drug used to take over a decade and cost upwards of $2.5 billion. Even then, around 90% of drug candidates failed before reaching patients. In 2026, AI is transforming that equation. Machine learning can now screen billions of molecular compounds in days, predict how proteins fold in three dimensions, identify patient populations most likely to respond to a specific treatment, and flag safety risks before a single human trial begins. The UK sits at the centre of this revolution, home to world-class pharmaceutical companies, elite research universities and NHS data sets that most countries can only dream of accessing. This guide explains how AI drug discovery works, who is leading the charge in the UK, and what it means for patients and investors.

What Is AI Drug Discovery?

Traditional drug discovery begins with a target — a protein or biological mechanism involved in a disease. Scientists then screen thousands of chemical compounds to find one that interacts with that target in a useful way. This hit-to-lead process was laborious and expensive, often taking four to five years just to identify a promising candidate. AI changes it by training machine learning models on vast databases of known compounds, biological structures and clinical outcomes, then using those models to predict which new compounds are likely to work — and which are likely to fail — before any laboratory work begins.

The result is a dramatic reduction in the early discovery phase. Insilico Medicine used AI to identify and design a novel drug candidate for idiopathic pulmonary fibrosis in just 18 months — a process that typically takes four to five years through conventional methods. The compound entered Phase 2 clinical trials in 2023 and reported promising early results. The same computational pipeline has since been applied to dozens of other disease areas, demonstrating that this is not a one-off success.

In the UK, the Alan Turing Institute and the Wellcome Sanger Institute have been central to building the foundational models and curated datasets that underpin AI drug discovery research. Their work sits alongside major commercial operations at companies like Exscientia, BenevolentAI and GSK’s dedicated AI research unit in Stevenage.

How Machine Learning Screens Billions of Compounds

The human body contains approximately 20,000 proteins. Each protein can interact with millions of potential drug molecules. The number of possible interactions is effectively infinite — far beyond what any human team could evaluate experimentally. AI makes this computationally tractable by learning patterns from known drug-protein interactions and extrapolating to novel cases.

Graph neural networks (GNNs) are particularly useful in this context. A molecule can be represented as a graph, where atoms are nodes and chemical bonds are edges. A GNN can process that structure and predict properties like binding affinity, toxicity and solubility, without running a physical experiment. A model trained on 100 million known molecules can evaluate the likely properties of a novel compound it has never seen before, in seconds.

Generative AI has added another powerful dimension. Rather than screening existing compounds from a library, generative models can design entirely new molecules from scratch, optimised for specific target properties. Exscientia, headquartered in Oxford, used this approach to generate a drug candidate for obsessive-compulsive disorder that entered clinical trials in under 12 months — roughly four times faster than conventional methods. The company has since partnered with Sumitomo Pharma and Bayer to scale the approach across their broader pipelines.

Virtual screening now handles the equivalent of what physical wet lab screening would take decades to test. Computation that would have required a supercomputer cluster in 2010 can run on a standard cloud server in 2026, thanks to both hardware improvements and far more efficient algorithms.

Protein Folding: AlphaFold’s Revolution

No AI breakthrough has affected drug discovery more profoundly than AlphaFold. Developed by DeepMind — the UK-based AI research company now part of Google — AlphaFold solved a 50-year-old challenge in biology: predicting the three-dimensional structure of a protein from its amino acid sequence alone. Protein structure determines function, and function determines how a protein can be targeted by drugs.

Until 2020, determining a protein’s structure required expensive, time-consuming laboratory techniques like X-ray crystallography or cryo-electron microscopy. A single protein structure could take years to resolve. AlphaFold2, released in 2021, predicted protein structures with accuracy comparable to experimental methods — and did so in minutes. It represented a genuine paradigm shift, widely regarded as one of the most significant scientific achievements of the century.

In July 2022, DeepMind and the European Bioinformatics Institute released predicted structures for virtually all 200 million known proteins in the public AlphaFold Protein Structure Database. Researchers worldwide can access this database for free. In 2024, AlphaFold 3 extended the model to predict how proteins interact with other molecules — DNA, RNA and small drug-like molecules — which is directly applicable to drug design. The tool has been cited in over 20,000 scientific publications as of early 2026.

The impact on UK research has been profound. Scientists at the MRC Laboratory of Molecular Biology in Cambridge have used AlphaFold predictions to identify new antibiotic candidates against drug-resistant bacteria. Researchers at Imperial College London used it to identify potential targets for a new class of cancer immunotherapy drugs that is now entering early-stage trials.

Clinical Trials: How AI Finds the Right Patients

Even when a promising drug compound is identified, clinical trials remain the most expensive and time-consuming part of the development process. A Phase 3 trial — testing effectiveness in large patient groups — typically takes three to five years and costs hundreds of millions of pounds. Failure rates remain stubbornly high. AI is beginning to change the economics and the odds.

Patient stratification is one of the most impactful improvements. Instead of enrolling anyone with a particular diagnosis, AI models can analyse genomic, clinical and lifestyle data to identify the specific subset of patients most likely to respond to a particular treatment. This approach reduces required trial size, cuts cost and duration, and improves the chances of seeing a statistically significant result — all while delivering better outcomes for the patients who participate.

The NHS has a structural advantage here that no other healthcare system can match. Its Electronic Health Records contain data on over 60 million patients going back decades. Unlike the fragmented US healthcare system, NHS data is relatively centralised and standardised. Federated learning — a privacy-preserving technique that allows AI models to train on data without it leaving hospital systems — allows researchers to access this resource without breaching patient confidentiality. The NHS AI Lab, launched by NHSX, has been central to enabling this kind of research.

Real-world evidence tools allow AI to analyse outcomes from patients already taking a medicine outside of controlled trial conditions. This can substantially shorten the evidence base required for regulatory approval. The Medicines and Healthcare products Regulatory Agency (MHRA) published a framework for incorporating real-world evidence into regulatory decisions in 2024.

UK Pharma and AI: Who Is Leading the Charge?

The UK is exceptionally well-positioned to lead AI drug discovery. It is home to two of the world’s largest pharmaceutical companies — AstraZeneca and GSK — both of which have made AI central to their research strategy. AstraZeneca’s AI Centre in Cambridge employs over 300 AI specialists and has committed to a $1.2 billion partnership with Recursion Pharmaceuticals to apply AI systematically across its drug pipeline. GSK created a standalone AI and machine learning group in 2022 and has since used it to identify over 40 novel drug targets across oncology, respiratory disease and immunology.

Exscientia, founded at the University of Dundee, is one of the world’s most prominent AI-first drug discovery companies. Its platform has generated multiple clinical candidates across oncology, psychiatry and immunology, with partnerships spanning Bristol-Myers Squibb, Bayer and Sanofi. The company listed on Nasdaq in 2021, reflecting significant international investor interest in UK-developed AI drug discovery capabilities.

BenevolentAI, based in London, uses knowledge graphs — structured databases of biological relationships — combined with machine learning to identify drug targets for diseases including ALS, Parkinson’s and inflammatory bowel disease. After financial restructuring in 2023, the company refocused on its core computational platform and partnership model. It represents the kind of specialised UK AI capability that has attracted sustained global pharmaceutical interest.

UK government support has been consistent. Innovate UK and UKRI invested over £300 million in AI and life sciences research between 2022 and 2025, as part of the Life Sciences Vision. The Medicines Discovery Catapult in Alderley Edge specifically helps smaller UK biotech companies access AI drug discovery tools they could not otherwise afford, lowering the barrier to entry for early-stage innovation.

The Limitations of AI in Drug Discovery

Despite the genuine progress, AI has not solved drug development. It has made the early discovery phase significantly faster and cheaper — but it has not meaningfully improved success rates in clinical trials, which remain stuck around 10%. Biology is still full of surprises that no computational model can fully anticipate.

Interpretability is a persistent challenge. Many AI models are black boxes — they produce predictions without explaining the reasoning behind them. In drug development, understanding the mechanism of action is critical for both scientific validity and regulatory approval. A model that predicts “this compound will bind to this target” without explaining how provides limited scientific value and makes it harder to optimise the compound further.

Data quality and bias are significant constraints. AI models learn from historical drug trial data, which is often incomplete, biased toward certain disease types and demographic groups, and stored in incompatible formats across different institutions. Models trained on biased data will produce biased predictions — potentially missing drug candidates that would work well in under-represented populations.

Off-target effects — unintended interactions between a drug molecule and proteins other than its intended target — remain difficult to predict reliably. Some of the most catastrophic drug failures resulted from off-target binding that no screening method anticipated. Modern AI models have become better at flagging this risk, but they are not infallible, and the consequences of a missed warning in human trials can be severe.

Regulatory Hurdles: The MHRA and AI-Developed Drugs

Bringing an AI-designed drug to market requires navigating regulatory frameworks that were designed long before AI existed. The UK’s MHRA has taken a notably proactive approach. It published the AI Roadmap for Medicines in 2025, setting out how AI-generated evidence will be evaluated in drug approval submissions and what standards AI systems must meet to be considered reliable.

The core regulatory principle is that AI-generated evidence must be validated against real-world experimental outcomes before it can substitute for traditional laboratory data. A model that predicts a compound will be safe must be shown to be consistently accurate across a broad range of known cases — and its failure modes must be clearly characterised — before its predictions are accepted as primary evidence in a submission.

A particularly complex challenge involves AI systems that continuously update as they process new patient data. A model that improves over time may behave differently from the version reviewed during initial regulatory approval. This creates genuine questions around version control, transparency and regulatory consistency. The MHRA is working through these questions collaboratively with developers, rather than imposing prescriptive requirements before the technology has matured.

International coordination is increasingly important. The US FDA, European Medicines Agency and MHRA are actively collaborating on harmonised standards for AI in drug development, to avoid a patchwork of incompatible national requirements that would slow global clinical programmes and create unnecessary duplication of regulatory effort.

What This Means for UK Patients and Investors

For UK patients, the near-term impact of AI drug discovery is most likely to be felt in oncology and rare diseases. These fields have historically been underserved by conventional research — cancer has too many variants and rare diseases too few patients to justify the cost of traditional large-scale trials. AI reduces the cost of exploring these spaces dramatically, making it economically viable to develop treatments for conditions that previously had no commercial development pathway.

For the NHS, AI drug discovery holds the promise of more targeted, personalised treatments — medicines designed for specific patient subgroups with better efficacy and fewer side effects, rather than broad population averages. The challenge is cost. Personalised medicines developed through AI tend to command premium pricing, and the NHS will face difficult decisions about which AI-developed drugs to fund through its technology appraisal process, just as it does with cell and gene therapies today.

For investors, AI drug discovery companies present a nuanced risk profile. The upside is substantial: a company that consistently accelerates drug development has enormous long-term commercial value. The downside is that even strong computational predictions frequently fail in clinical trials, and biology does not always cooperate with models. Careful due diligence should focus on the diversity of the pipeline, the quality of the underlying data assets, and the strength of partnership agreements with established pharmaceutical companies who can validate and commercialise the discoveries.

This article is for educational purposes only and does not constitute financial advice.