An AI search for cocaine addiction medicines led experts to ketamine, a drug already being studied

The short version

Researchers used an AI system to rank approved medicines that might help cocaine use disorder, which has no FDA-approved medication. An expert panel picked ketamine from the list. In health records, people with cocaine use disorder who had been prescribed ketamine were more likely to later have a remission diagnosis than similar people given other drugs. The link comes from past records, not a trial. Remission diagnoses were uncommon, and the authors say trials are needed.

What they did

The work was part of a National Institute on Drug Abuse Clinical Trials Network project (CTN-0114). It had four steps.

First, an AI system called KG-Predict ranked 1,430 FDA-approved drugs. It uses a knowledge graph, a map of known links between drugs, genes, diseases, and symptoms, built from public databases and medical text. Second, a 7-member advisory committee of addiction researchers reviewed the top 35. Ketamine ranked sixth on the AI's list and was the only drug every committee member recommended.

Third, the team searched TriNetX, a network of health records from about 90 million patients. They found 379,409 people diagnosed with cocaine use disorder between January 2007 and June 2022, of whom 16,754 had been prescribed ketamine. They compared people given ketamine within a year of diagnosis to matched people given something else. Fourth, they looked at which genes and biological pathways ketamine and cocaine use disorder share.

What it found

In the anesthesia comparison (3,871 people in each group), ketamine was linked to about twice the rate of remission diagnoses (hazard ratio 1.98, 95% CI 1.42 to 2.78). In the depression comparison (3,955 in each group), against antidepressants or midazolam, the hazard ratio was 4.39 (95% CI 2.89 to 6.68).

The authors repeated the analysis in two groups that did not overlap: people given ketamine for anesthesia who never had depression, and people given it for depression who never had anesthesia. The link held in both (hazard ratios 2.23 and 3.37), with wider confidence intervals.

The absolute numbers were small. The authors report remission rates of 1% to 3% among ketamine-treated patients. Roughly doubling or quadrupling a small rate still leaves a small rate. The authors still call the results "a convincing signal for potential clinical efficacy." We read them more cautiously, for the reasons below.

What it does not show

This study does not show that ketamine treats cocaine use disorder. No one was given ketamine to test it. The comparison used past records, and the authors state that matching cannot rule out differences between groups that were not measured.

Remission here means a diagnosis code for cocaine use disorder in remission. It is not a urine test or a measured drop in use. The records did not show dose, how long people took ketamine, or whether they took it as prescribed. Most people in the depression group got a single prescription.

The authors are candid about the AI's part. Ketamine had already been studied for cocaine use disorder in small trials, so the AI did not find a new candidate among the 35 the experts reviewed. Most of the other candidates were rated low, largely over concerns that patients would not keep taking them. The authors say the study shows why AI output needs expert vetting.

The analysis was not pre-registered, and the authors call the results exploratory. The work was funded by the National Institute on Drug Abuse and the Clinical and Translational Science Collaborative of Cleveland. One co-author is the NIDA scientific officer for the project. The authors declared no conflicts of interest. The funding statement names no drug company.

Why it matters

For people in recovery, the takeaway is narrow. This is a lead for researchers, not a treatment. Nothing here supports seeking ketamine for cocaine use outside a clinical trial.

For anyone following AI in addiction research, this paper is a good picture of how the work actually goes. The AI produced a ranked list, human experts did the choosing, and health records supplied a signal. None of those steps tested a treatment. The authors say additional clinical trials are warranted. Until they are done, ketamine for cocaine use disorder is a hypothesis.