AI & Science
For a century, finding a new drug was a game of chance played at enormous cost. Artificial intelligence is changing the odds -- and the first results from clinical trials are beginning to arrive.
In 2022, a small biotech company called Insilico Medicine sent a drug candidate into Phase I clinical trials. The molecule had been designed entirely by artificial intelligence -- from the identification of a disease target to the suggestion of a chemical compound that might hit it -- in less than 18 months. The same process, done conventionally, would have taken five years or more. The drug, ISM001-055, subsequently produced positive Phase IIa results in idiopathic pulmonary fibrosis, a fatal lung disease with few treatment options. It was one of the first AI-designed molecules to show clinical promise. It would not be the last.
Modern drug development is one of the most expensive and failure-prone enterprises in science. Most drug candidates that enter development never reach patients. A widely cited analysis from the Tufts Center for the Study of Drug Development estimated the average cost of bringing a single new drug to market -- accounting for failures along the way -- at approximately $2.6 billion. The process takes, on average, ten to fifteen years from initial discovery to regulatory approval.
The attrition is staggering. Roughly nine out of ten drug candidates fail before reaching patients, the majority in Phase II and III trials, after substantial capital and years of work have already been committed. Failures happen for two main reasons: the molecule does not work well enough (lack of efficacy), or it produces harmful effects in humans that animal models did not predict (toxicity). Both failures reflect limitations in how well researchers can predict molecular behaviour before running expensive human trials.
The cost is not just financial. Diseases without profitable drug markets -- rare genetic conditions, neglected tropical diseases, drug-resistant infections -- go without treatments precisely because the economics of conventional drug development make unprofitable targets unattractive. A faster, cheaper pipeline would change that calculation.
For decades, the industry response to these constraints was incremental: better analytical chemistry, more automated screening, improved animal models. The pipeline became slightly more efficient, but not fundamentally different. What AI offers is something structurally different: the ability to reason about molecular space at a scale and speed that no team of human chemists can match.
The challenge of drug discovery, viewed mathematically, is a search problem of bewildering scale. Estimates of drug-like chemical space -- molecules that could in principle be synthesised and tested as drugs -- range from 1023 to 1060 distinct compounds. No physical library of compounds can hold more than a few million. Conventional high-throughput screening, which tests large collections of existing compounds against a disease target, is sampling a tiny fraction of the possible space. Most of the ocean of potential drugs has never been explored.
A drug candidate is a molecule that has been identified as having a desired biological effect -- typically binding to a target protein associated with disease -- and has enough other properties (stability, ability to reach its target in the body, low toxicity) to warrant entering formal clinical testing. The challenge is that these requirements are often in tension: a molecule that binds its target well may be metabolised too quickly, or may also bind other proteins and cause side effects. Drug discovery is the art of finding molecules that thread all of these needles simultaneously.
AI approaches the problem differently. Instead of physically testing compounds one by one, machine learning models can be trained on existing data -- millions of known drug-target interactions, protein structures, toxicity measurements -- and then used to predict which molecules in unexplored chemical space are most likely to have the desired properties. Generative AI models can go further, designing entirely new molecules from scratch rather than selecting from existing libraries.
A pivotal moment came in 2020 when DeepMind's AlphaFold2 demonstrated that deep learning could predict three-dimensional protein structures from amino acid sequences with remarkable accuracy. The system subsequently produced predicted structures for approximately 200 million proteins, making that database freely available to researchers worldwide. Knowing a target protein's structure -- how its binding sites are shaped, where small molecules might attach -- is fundamental to designing drugs that interact with it. AlphaFold changed the information landscape of drug discovery overnight.
AI compresses discovery-phase timelines substantially; clinical trial phases remain similar. Timelines are indicative estimates based on published literature.
The practical tools vary by company and stage of development. Virtual screening uses AI models to rank vast libraries of compounds by predicted binding affinity. De novo design generates new molecular structures optimised for multiple properties at once -- efficacy, selectivity, metabolic stability, synthetic accessibility. Toxicity prediction models flag candidates likely to produce adverse effects before any molecule is synthesised. Together, these tools allow researchers to do computationally in days what would have taken laboratory years.
As of the end of 2024, more than 75 AI-derived molecules had reached clinical stages -- a small but growing cohort whose results are beginning to provide real data about whether AI-designed drugs perform differently from conventionally discovered ones. The early signal is cautiously encouraging, though the sample size remains limited.
A 2024 analysis published in Drug Discovery Today found that AI-native biotechs had achieved Phase I success rates of approximately 80 to 90 percent, substantially above the historical industry average of around 40 to 65 percent. Phase II results, on a smaller sample, showed roughly 40 percent success -- above the historical industry average of 29 percent. These figures should be interpreted carefully: the AI-focused companies represent a self-selected group, the molecules are early, and selection bias may inflate reported results. But they are not trivial numbers.
Beyond Insilico Medicine, several other AI-designed molecules have progressed. Schrödinger's zasocitinib (TAK-279), a TYK2 inhibitor for autoimmune diseases, advanced into Phase III clinical trials -- one of the furthest-progressed AI-originated molecules in the pipeline. Exscientia, Recursion Pharmaceuticals, and Absci are among the companies with AI-designed candidates in clinical stages across oncology, immunology, and rare disease.
"AI does not eliminate the risk in drug development. It compresses the early stages where human intuition has always been the weakest, and focuses experimental effort where the probability of success is highest."
-- Lisa PedrosaWhat is less clear is whether AI improves Phase III success rates -- the stage where most drug development costs are concentrated. Phase III failures typically reflect disease biology that was not fully understood at the outset, patient population heterogeneity, and regulatory requirements that no amount of early computational elegance can fully anticipate. The hardest parts of drug development may not be where AI's advantages are strongest.
On April 14, 2026, Novo Nordisk -- the Danish pharmaceutical giant whose GLP-1 obesity drugs generated more revenue than the GDP of many countries -- announced a strategic partnership with OpenAI. The companies stated that the collaboration would integrate advanced AI across Novo Nordisk's research and development, manufacturing, and commercial operations, with pilot programs launching immediately and full integration targeted by the end of 2026.
The announcement was notable less for its technical novelty than for its institutional signal. Novo Nordisk is not a startup running AI experiments in a corner of its R&D department. It is one of the most commercially successful pharmaceutical companies on earth. When a company of that scale commits to enterprise-wide AI integration -- with explicit language about analysing datasets "at a scale that was previously impossible" and identifying drug candidates at speed -- it suggests an industry at a genuine inflection point rather than a speculative one.
The partnership covers not only drug discovery but also the company's supply chain, workforce training, and operational efficiency. This broader framing reflects a maturation in how large pharmaceutical companies think about AI: not as a single laboratory tool, but as an infrastructure layer that touches every part of how a drug goes from idea to patient.
The promise of AI in drug discovery includes the possibility that faster, cheaper development could open up disease areas that were previously uneconomical. Rare diseases affecting small patient populations have historically been neglected because the cost of conventional development outweighs the commercial return. If AI can cut discovery-phase costs by an order of magnitude, the economics of rare disease drug development changes. Whether this theoretical benefit translates into actual drug development for neglected diseases -- rather than simply accelerating development of drugs in competitive commercial markets -- depends on decisions made by companies and regulators, not algorithms.
The longer arc of this technology is genuinely difficult to predict. AI has already changed what is computationally possible in early-phase drug discovery. Whether it changes drug development's overall success rate in a statistically unambiguous way will require another decade of clinical data. What is clear is that the pipeline is being rebuilt around different tools, and that the industry's largest players have committed to that rebuilding in ways that are no longer reversible.
For patients waiting for treatments that do not yet exist, the relevant question is not whether AI drug discovery is real -- the clinical trials have settled that -- but how quickly the pipeline translates computational progress into clinical outcomes. The first AI-designed drug to reach full regulatory approval has not yet happened. When it does, the debate about whether AI changes drug development will effectively be over.
© 2026 Lisa Pedrosa · lisapedrosa.com
All articles cited to primary institutional or peer-reviewed sources
Buy me a coffee