A review published on September 10 sets out how artificial intelligence could help pharmaceutical researchers recruit trial participants, organize evidence and decide sooner whether a treatment deserves further testing. It also stresses that these tools need validation and human oversight before their promised savings can be trusted.
The Nature Reviews Bioengineering review, led by Marc Raynaud and colleagues, proposes a framework for clinical trials. It is an assessment of methods and examples, not a new experiment demonstrating an industry-wide reduction in development costs.
AI is already appearing well beyond the laboratory. The US Food and Drug Administration’s Center for Drug Evaluation and Research reports experience with more than 500 submissions containing AI components between 2016 and 2023. The applications span preclinical research, human trials, manufacturing and activities after medicines reach the market.
Those are submissions containing AI, not 500 approved AI-designed drugs. The business question is whether the technology improves enough decisions to justify the cost of introducing and maintaining it.
Finding molecules is only the beginning
In drug discovery, machine learning, which identifies patterns from training data, can help researchers choose biological targets and assess compounds. A target might be a protein involved in a disease. Generative models can also propose molecular structures for scientists to investigate.
Computer predictions can narrow the list of experiments to perform. A molecule must still be made and tested, and a promising laboratory result may not translate into a treatment that works safely in people.
An August 7 perspective in Nature Reviews Drug Discovery, led by Andreas Bender, found that evidence of AI’s clinically relevant impact remained “so far, disappointingly limited.”
The authors argue that evaluations should ask whether a model helps people make better drug-development decisions. An impressive score on a prepared test dataset offers limited reassurance if the tool performs poorly on the unfamiliar compounds or patients encountered in practice.
A real trial shows progress and remaining uncertainty
One example of clinical progress is rentosertib, developed using generative AI to help identify both a disease-related target and a compound. A phase 2a trial published in Nature Medicine in June 2025 enrolled 71 patients with idiopathic pulmonary fibrosis, a disease that progressively scars the lungs.
Participants were randomly assigned to three dosing groups or a placebo for 12 weeks. The primary measure concerned adverse events arising during treatment. A secondary lung-function measure improved on average in the highest-dose group.
The findings require caution. The trial was small and short, treatment-related adverse events were more common in the drug groups, and liver problems caused some participants to stop treatment. Insilico Medicine sponsored the study, and several authors were its employees.
The study tested the medicine, not whether an AI research process outperformed a conventional one. Its results supported further investigation; they did not establish long-term benefit or prove that AI makes drug development more successful across the industry.
Helping trials run more efficiently
The September review discusses matching medical records to trial eligibility rules and organizing information from different sources. Such tools could reduce manual work while clinicians retain responsibility for enrollment decisions.
It also examines digital twins, computational representations that simulate aspects of a patient or organ, and comparison groups assembled from existing patient data. These approaches depend on reliable data and assumptions. Simulated outcomes cannot automatically substitute for evidence from randomized trials.
For a company funding several research programs, learning earlier that one is unlikely to work could free money and staff for others. But abandoning a promising treatment because of an unreliable prediction would destroy that advantage. Faster decisions have economic value only when the evidence behind them is dependable.
Manufacturing needs continuous checks
A 2026 review in the Journal of Pharmaceutical Sciences describes AI applications in manufacturing, including detecting unusual process behavior, estimating product characteristics from sensor readings and inspecting injectable medicines with computer vision.
A model might flag a developing production problem before a batch is finished. Computer vision can examine images for defects, while other models analyze several measurements together to identify patterns that individual alarms might miss.
The review places these applications within Good Manufacturing Practice, the quality requirements governing medicine production. Manufacturers need documented evidence that a model works for its intended task, traceable data and monitoring as operating conditions change.
Quality control also continues after production. As we reported in our coverage of the pharmaceutical cold chain, temperature-sensitive medicines require controlled storage and transport. A more efficient discovery process leaves those obligations intact.
Regulators focus on the job each model performs
In January 2026, the FDA and European Medicines Agency published 10 principles for good AI practice in drug development. They cover defined uses, data governance, multidisciplinary expertise, performance assessment and management throughout a model’s life cycle.
The principles call for validation and oversight proportionate to risk. For example, helping staff organize documents raises different questions from producing evidence that influences whether a medicine should be authorized. Companies need to specify what a model does, which data it relies on and how errors will be detected.
The regulators describe the principles as a foundation for further good practice, not blanket approval of AI tools. Medicines still have to demonstrate quality, safety and effectiveness, with benefits that outweigh their risks. An algorithm’s involvement does not lower that standard.
