On a screen, a molecule can look ideal. It fits its target, shows an encouraging profile and, according to the algorithm, deserves to be made. Inside a patient’s body, the story becomes more complicated: insufficient absorption, adverse effects, a disease more heterogeneous than expected. The real test of drugs developed using artificial intelligence begins where the computational demonstration ends. The paths taken by Insilico Medicine and Recursion shed light on this gap between discovery speed and medical evidence. Looking ahead to September 2026, results already published offer grounds for hope, but certainly no license to draw conclusions.
Faster discovery does not mean a faster cure
Developing a drug involves a succession of bets. Researchers must identify a relevant biological mechanism, find a substance capable of modifying it, verify that it reaches the right tissue and determine a tolerable dose. They must then demonstrate that this intervention actually improves health. AI can contribute at every stage, but the evidence generated is not interchangeable.
A model that performs well at predicting how a molecule binds to a protein cannot necessarily predict whether a hospitalization will be prevented. A successful experiment on cells reproduces neither aging nor concomitant treatments, nor the full immune complexity of a patient. Accelerating candidate selection shortens the research timeline; it does not automatically validate the therapeutic hypothesis.
The terminology sometimes adds to the confusion. “AI-designed drug” can refer to a chemical structure generated by a model, a target identified in biological data or an existing molecule repurposed for a new indication. These approaches have different merits. Grouping them under a single label makes communication easier, but does less for scientific evaluation.
Insilico: a flagship candidate takes on pulmonary fibrosis
Insilico Medicine illustrates the approach that combines target discovery with molecule generation. Its candidate rentosertib, previously known as INS018_055, targets TNIK, a kinase involved in mechanisms associated with fibrosis. The company developed it for idiopathic pulmonary fibrosis, a progressive disease in which lung tissue deteriorates and breathing becomes difficult.
The program is being watched particularly closely because it gives concrete form to a long-abstract promise: taking an algorithm-generated proposal from the laboratory into humans. AI helped select the target and then design the candidate. That represents a development success, without in itself demonstrating that the treatment will produce a lasting change in the course of the disease.
Phase IIa results published in 2025 in Nature Medicine provided an encouraging initial clinical signal. This small, short-duration study primarily examined safety and tolerability, while also exploring measures including lung function. The findings justify further research; they do not replace confirmatory evidence from a larger population.
In this disease, a favorable change in forced vital capacity—the volume of air exhaled during a breathing maneuver—deserves attention. But researchers need to establish whether the effect persists, whether it holds up under robust statistical analysis and how it fits alongside available treatments. Quality of life, exacerbations and long-term tolerability also matter. A small study cannot provide firm answers to all these questions.
Recursion: mapping biology rather than simply designing molecules
Recursion is taking a different path, even as the approaches converge. The company combines automated experimentation, cellular imaging and machine learning. Its ambition is to observe how cells respond to genetic or chemical perturbations, then identify relationships that could point toward a therapeutic avenue. Here, the algorithm helps read a vast experimental map of biology.
The merger with Exscientia, completed in 2024, combined this approach with AI-assisted molecular design expertise. The business objective is understandable: to connect biological discovery, chemistry and clinical development more closely. But bringing platforms together does not yet prove that the drugs they produce will be more successful in trials.
The case of REC-994 also highlights why each program must be examined separately. Evaluated for cerebral cavernous malformations, this molecule is not an entirely new creation of generative AI. Recursion announced in 2024 that its phase II study had met its primary safety and tolerability endpoint, with exploratory findings drawn notably from imaging. Meeting a safety endpoint is not equivalent to demonstrating clinical efficacy.
For affected patients, the decisive question remains the consequences of these lesions: symptoms, disability or neurological events. A radiological change can guide research without, on its own, providing sufficient evidence of benefit. Further development of such programs must therefore be judged on published results, rather than inferred from the platform’s claimed performance.
Why trials remain essential
A well-designed clinical trial establishes a comparison that individual impressions cannot provide. Randomization limits baseline differences between groups. Blinding reduces certain influences on assessment. A predefined primary endpoint prevents researchers from choosing the most flattering result from dozens of measurements after the fact.
Three questions can help make sense of announcements:
- What was the study designed to demonstrate? Tolerability, an appropriate dose and a therapeutic benefit are not the same objective.
- How robust is the result? Sample size, comparator group, duration and statistical uncertainty determine its significance.
- What changes for the patient? An encouraging biomarker does not always translate into a noticeable or lasting improvement.
AI can help with recruitment, identifying certain patient profiles or analyzing complex data. It can also introduce bias if its recommendations rely on unrepresentative populations. These tools must be validated for their intended use. They do not remove the need for a rigorous protocol or careful monitoring of adverse effects.
The true assessment will span multiple programs
A first approval would be a major milestone, not universal proof of AI’s superiority. Conversely, the failure of one candidate would not discredit every method. Assessing their contribution will require comparing portfolios of projects: time saved, expenditure, candidate quality and the likelihood of progressing through clinical stages.
That assessment will need to include dead ends. A platform can create value by abandoning an unpromising avenue earlier, before an expensive trial. But those decisions must be made visible. The discovery timeline so often highlighted tells one part of the story; the full cost of an effective drug tells the other.
What happens next? The most credible outlook for September 2026 and beyond is not one in which trials have become unnecessary, but one of better-selected candidates and better-prepared studies. Insilico and Recursion will need to turn their technological advantages into reproducible benefits for patients. AI will truly have transformed drug development when it helps produce, more efficiently, treatments whose value stands up to clinical comparison.