From molecular design to trial recruitment, artificial intelligence is influencing the choices behind drug development. European researchers and companies must distinguish faster selection from demonstrated therapeutic benefit.
A drug discovery algorithm and a clinical trial recruitment system appear to solve different problems. One searches chemical possibilities; the other searches patient records. Their common purpose is to improve a consequential decision about what happens next. Which molecule deserves synthesis, which biological target deserves investment, and which patient should be considered for a study?
For Europe’s pharmaceutical industry, this is a more useful way to assess artificial intelligence than counting compounds described as AI-discovered. Each application changes a particular decision and its associated costs. A better molecular prediction can reduce an unproductive laboratory experiment. A better trial search can help a research nurse find an overlooked option. Neither result automatically establishes that a medicine improves health.
The emerging evidence combines genuinely new scientific capabilities with familiar development constraints. It also reveals why drug developers, research hospitals and technology suppliers need different measures of success.
A designed molecule reaches a clinical test
Rentosertib offers a concrete example of the journey. In a phase 2a study published in Nature Medicine in June 2025, researchers tested the TNIK inhibitor in 71 people with idiopathic pulmonary fibrosis. Generative AI had contributed to identifying the target and designing the compound. Participants in China received one of three dosing regimens or placebo for 12 weeks. The primary endpoint concerned treatment-emergent adverse events.
The highest daily dose group showed a mean increase in forced vital capacity of 98.4 millilitres, compared with a decline of 20.3 millilitres on placebo. This lung function measure was a secondary endpoint. Treatment-related adverse events were more frequent with rentosertib, and liver toxicity contributed to discontinuations. The small, short trial supported further investigation; it did not establish durable benefit or superiority to established treatment.
The industrial implication is precise. AI helped produce a testable therapeutic hypothesis. Investment decisions beyond that point must still account for dose selection, drug interactions, tolerability and the probability that an early signal survives a larger trial. The compound’s computational origin cannot answer those questions.
Structural prediction changes laboratory choices
AlphaFold 3 expanded the types of molecular interactions that researchers could predict. The 2024 Nature paper described joint structure prediction involving proteins, nucleic acids, small molecules and other components, with improvements on several benchmark tasks. These are capabilities relevant to exploring how a potential medicine interacts with a biological target. They are not measurements of clinical efficacy.
For a discovery team, the valuable output may be a shorter and better justified experimental shortlist. A predicted interaction can help researchers decide which chemical modifications to investigate, which assay to run or which structural question requires direct measurement. The practical comparison is therefore between alternative research workflows, including the experiments they avoid and the useful candidates they preserve.
That comparison should include failures. Reporting only the molecules that progress makes it difficult to judge whether a platform improves selection across a portfolio. A commercially meaningful assessment would track comparable programmes from their starting hypotheses, recording laboratory expenditure, unsuccessful predictions and time to a decision to continue or stop. This is a proposed assessment approach, rather than a result established by the structural prediction study.
Repurposing provides a different route to benefit
An existing medicine can start with a different evidence base. In February 2020, researchers associated with BenevolentAI and Imperial College London proposed baricitinib as a potential COVID-19 treatment after using an AI-supported search of existing biomedical knowledge. This was a repurposing hypothesis for an established drug, not the invention of a new molecule.
The subsequent RECOVERY randomised trial supplied clinical evidence. Among 8,156 participants allocated to baricitinib or usual care, 514 of 4,148 in the baricitinib group died within 28 days, compared with 546 of 4,008 receiving usual care. The published analysis found a reduction in mortality.
This history separates two contributions often compressed into a single success story. Computational reasoning helped prioritise an intervention, while the trial tested its effect in patients. The result does not quantify how much faster the overall programme became because of AI. For European translational research, it nevertheless shows the importance of connecting hypothesis generation to organisations capable of conducting decisive studies.
Trial matching has several different endpoints
Language models can also assist the detailed reading needed to match people with trial criteria. TrialGPT, reported in Nature Communications in 2024, combined trial retrieval, assessment of individual eligibility criteria and ranking. Its small user study involved two doctors, six oncology trials and six semi-synthetic patient vignettes derived from anonymised referrals. Screening took approximately 42.6% less time with assistance.
That is a screening efficiency result in a constructed task. It does not show a 42.6% improvement in recruitment, trial completion or patient survival. Moving from a possible match to participation involves checking eligibility, discussing the study, obtaining consent and determining whether a suitable place is available.
For a hospital or contract research organisation, those distinctions change what to buy. A search service may be worthwhile if it reduces the burden of reviewing unsuitable trials. A recruitment service claiming to increase enrolment should be assessed on actual enrolment, with reasons for failed referrals recorded. Paying for a list of plausible matches and paying for an operational recruitment improvement are different propositions.
European trial searches depend on current local information
A prospective evaluation at Centre Léon Bérard in France examined four available matching tools in 157 consecutive molecular tumour board patients during ten weeks in 2024. The study, published in January 2025, assessed DigitalECMT, Klineo, ScreenAct and Trialling. Across the tools, mean precision was 0.33 and mean sensitivity was 0.32. These particular measures were calculated in the subset with at least one eligible trial across the tools.
Errors frequently concerned the precise molecular alteration required. Recruitment status also caused false matches. A separate retrospective experiment added a local language model to rerank results for 34 patients with suitable molecular reports; its ranking improvement was modest. The prospective comparison should not be described as a trial of four large language models.
This points to an operational asset that is easy to overlook: reliable information about what a site can actually offer. A sophisticated interpretation of a genomic report cannot make a closed trial accessible. Regional research networks therefore have a practical role alongside software developers, maintaining referral arrangements and accurate study information.
Partnership contracts reveal where value is expected
Commercial arrangements show that discovery platforms can monetise research decisions before a medicine reaches patients. In January 2024, Isomorphic Labs announced small-molecule collaborations with Lilly and Novartis, including upfront payments of $45m and $37.5m respectively. Its stated potential value of nearly $3bn depended on further milestones and excluded royalties. That headline amount was not revenue already received.
These structures distribute development risk across stages. An upfront payment supports access to capabilities and research work; later payments depend on agreed events. By contrast, a hospital trial-matching subscription is purchased to support a recurring service. The two businesses may use related AI techniques, but their customers, time horizons and proof of value differ.
Europe’s medicines regulator addresses AI across this broader lifecycle. The European Medicines Agency’s adopted 2024 reflection paper sets out a risk-based approach, with greater scrutiny where an application can affect patient safety or the reliability of evidence used in regulatory decisions. It is a scientific reflection paper, rather than a marketing authorisation for a platform.
For smaller biotechnology companies, the choice also concerns ownership of the resulting knowledge. Commissioning a defined discovery project, licensing software and sharing downstream milestones place different responsibilities on the buyer. Contracts should make clear who can reuse experimental results, whether model updates change the service and what information remains available if a partnership ends. These are commercial design choices, not evidence that one model is universally preferable.
The strongest commercial proposition is consequently specific: a better target decision, a more informative experiment or a workable referral. Connecting those improvements to successful clinical development requires measuring each transition, including the decisions to stop. That is where Europe can turn computational progress into a more productive research system.
