Do AI drug repurposing systems have enough prospective clinical evidence?

What clinical evidence exists for AI drug repurposing systems?

The strongest evidence comes from retrospective analyses of electronic health records (EHRs) that test AI predictions against real patient outcomes. In one study, an AI knowledge graph identified ketamine as a candidate for amphetamine-type stimulant use disorder; a subsequent EHR analysis of over 3,600 patients found that those receiving ketamine had a 58% higher chance of remission compared to those receiving other anesthetics (hazard ratio 1.58) [1]. Another study used ChatGPT to rank repurposing candidates for Alzheimer's disease, then tested the top ten in two large clinical datasets (Vanderbilt and All of Us); metformin, simvastatin, and losartan were associated with lower Alzheimer's risk in a meta-analysis [2]. A third study applied machine learning to a database of over 12,000 patients with monoclonal gammopathy of undetermined significance (MGUS) and found that several drug classes—including multivitamins, statins, and beta-blockers—were associated with lower progression risk [3]. These studies share a common design: AI generates a hypothesis, then retrospective EHR data is used to check for a signal. This is a step beyond purely computational prediction, but it is not the same as a prospective clinical trial.

How strong is this evidence compared to traditional clinical trials?

The evidence is promising but has important limitations. All three clinical validation studies [1][2][3] are retrospective and observational, meaning they can show association but not causation. For example, the ketamine study [1] found a 58% higher remission rate, but patients who received ketamine may have differed in other ways (e.g., severity of illness) that influenced the outcome. Similarly, the Alzheimer's study [2] found lower risk with metformin, simvastatin, and losartan, but these drugs are commonly used for diabetes, high cholesterol, and hypertension—conditions that themselves affect Alzheimer's risk. The MGUS study [3] used a well-curated database and robust statistical methods (concordance index of 0.883, indicating good model fit), but the findings are still correlational. None of these studies are randomized controlled trials (RCTs), which are the gold standard for proving a drug works. The authors of the MGUS study explicitly state their work 'could inform subsequent prospective studies' [3], acknowledging that prospective trials are the next necessary step.

What about AI models that haven't been tested on patients yet?

Many AI repurposing systems are evaluated only on computational benchmarks, not on clinical data. For instance, a new foundation model called TxGNN was trained on a medical knowledge graph and outperformed eight other methods by 49.2% in predicting drug indications and 35.1% in predicting contraindications under zero-shot conditions (where no drug exists for a disease) [4]. The model's predictions also aligned well with off-label prescriptions in a large healthcare system. However, this is still a computational validation—the model has not been tested in a clinical setting. Another study used 3D-QSAR (a computational chemistry method) to predict which FDA-approved drugs might inhibit the protein sclerostin for bone disease; the model had strong statistical fit (r² = 0.969), but no patient data was involved [5]. These models are valuable for generating hypotheses, but they do not provide clinical evidence. The gap between a high-performing computational model and a proven treatment remains large.

About These Sources

This answer is built on 5 peer-reviewed studies — published from 2023 to 2025, 4 from 2024 or later, 3 in Q1 journals, collectively cited 183 times — selected as the most relevant from 5 studies that passed quality screening, drawn from 61 papers retrieved from a database of over 500 million.

Sources used in this answer

Related research questions