Repurposing ketamine to treat cocaine use disorder: Integration of artificial intelligence-based prediction, expert evaluation, clinical corroboration, and mechanism of action analyses

Cocaine use disorder (CUD) affects over 1.3 million adults in the United States and is associated with substantial morbidity and mortality, yet no FDA-approved pharmacotherapies exist. Standard care relies on psychosocial interventions, which are limited in effectiveness and reach. Drug repurposing offers a faster, less costly route to identify candidate medications from already approved drugs, and computational approaches such as network-, structure- and AI-based methods can generate large numbers of repurposing signals. However, many candidates flagged by such approaches fail in clinical testing, so additional vetting using clinical data and expert judgement is desirable. Gao and colleagues describe an integrated repurposing strategy that combines an AI-driven knowledge-graph prediction system (KG-Predict), advisory committee review, retrospective electronic health record (EHR) corroboration, and genetic/pathway analyses. The paper reports the outcome of applying this pipeline to CUD: ketamine emerged as a top candidate after AI ranking and unanimous expert recommendation, and the investigators then tested whether ketamine exposure in routine clinical care was associated with subsequent CUD remission and examined potential molecular pathways linking ketamine to CUD biology.

The study followed a multi-step pipeline. First, KG-Predict — a knowledge graph system integrating drugs, genes, diseases and phenotype annotations from multiple public databases — was used to prioritise candidate drugs from 1,430 FDA-approved compounds given a list of CUD-associated genes. CUD genes were sourced from DisGeNet and the literature; the final gene list included genes such as DRD2, SLC6A3, BDNF and GABRA2. The top 35 candidates from KG-Predict were then reviewed by a seven-member CTN-0114 advisory committee composed of experts in data science, addiction psychiatry and addiction trials; committee members provided recommendations via REDCap surveys and the investigative team selected ketamine for further evaluation based on consensus. Clinical corroboration used TriNetX, a federated EHR network covering approximately 90 million de‑identified patients. The investigators identified patients with a CUD diagnosis from January 2007 to June 2022 and extracted medication exposure (ketamine and comparator anesthetics, and ketamine and comparator antidepressants/midazolam). Three retrospective cohort comparisons were constructed: (1) CUD patients prescribed ketamine for anaesthesia versus matched CUD patients prescribed other anaesthetics; (2) CUD patients with unipolar/major depression prescribed ketamine versus matched CUD patients with depression prescribed antidepressants or midazolam; and (3) two non-overlapping cohorts isolating anaesthesia-only and depression-only ketamine exposures. The index event was the date of drug prescription, and patients who had previously remitted from CUD or whose prescription occurred more than 1 year after initial CUD diagnosis were excluded. To reduce confounding, the TriNetX built-in 1:1 propensity score matching (nearest-neighbour greedy algorithm, caliper 0.1 standardised mean differences) was used to balance covariates between exposure and control groups; the specific covariates are reported in the supporting information. Time-to-event comparisons for the primary outcome — a recorded diagnosis of CUD remission within 1 year of prescription — were analysed using Cox proportional hazards models, with hazard ratios (HR) and 95% confidence intervals (CI) reported and the proportional hazards assumption tested via the generalized Schoenfeld approach. The authors note the analysis was not pre-registered and should be considered exploratory. For mechanism-of-action work, the STITCH database was queried to identify protein-coding genes associated with ketamine (cut-off score median 0.5). KEGG pathway enrichment was performed for CUD-associated genes and for ketamine-associated genes; pathways with P < 0.01 were retained and intersected to identify shared pathways implicating both ketamine and CUD.

KG-Predict produced a ranked list of candidate repurposing drugs for CUD; the top 10 included aripiprazole, ketamine and quetiapine, three of which have prior clinical trial evidence in CUD. From the top 35 candidates provided to the advisory committee, ketamine ranked sixth and was the only drug unanimously recommended for EHR analysis. EHR analyses identified 379,409 patients with CUD in TriNetX from 2007–2022, of whom 16,754 had been prescribed ketamine. After propensity score matching, the anaesthesia cohort comprised 7,742 matched CUD patients (3,871 ketamine-exposed and 3,871 anesthetic-controlled). Patients who received ketamine for anaesthesia had a higher hazard of CUD remission within 1 year compared with matched patients receiving other anaesthetics (HR = 1.98, 95% CI = 1.42–2.78). Stratified analyses produced elevated HRs for both sexes: an HR of 2.32 (95% CI = 1.47–3.65) for women and an HR reported for men where the extracted text does not clearly report the complete 95% CI. Stratification by race also showed higher remission rates with ketamine: HR = 1.71 (95% CI = 1.09–2.68) for White patients and HR = 2.12 (95% CI = 1.24–3.63) for Black patients; the investigators report no significant gender or race disparities in ketamine-associated remission. In the depression cohort, after matching (7,910 patients; 3,955 ketamine-exposed and 3,955 antidepressant-controlled), ketamine exposure was likewise associated with greater CUD remission. The extracted text gives a confidence interval (95% CI = 2.89–6.68) but does not clearly provide the corresponding point estimate for the hazard ratio in the available extraction. Stratified analyses by gender and race showed similar findings without significant disparities. To address overlap between the anaesthesia and depression samples (approximately 25%), two non-overlapping cohorts were analysed. In the anaesthesia-only cohort (patients never diagnosed with depression), ketamine exposure was associated with higher remission (HR = 2.23, 95% CI = 1.02–4.91). In the depression-only cohort (patients who never received anaesthesia), ketamine exposure for depression was associated with higher remission (HR = 3.37, 95% CI = 1.45–7.83). At the genetic level, 154 genes were associated with ketamine in STITCH at the chosen threshold, of which 10 overlap with CUD-implicated genes: BDNF, CNR1, DRD2, GABRA2, GABRB3, GAD1, OPRK1, OPRM1, SLC6A3 and SLC6A4. KEGG enrichment identified 13 pathways significantly associated with the 24 CUD-associated genes, notably neuroactive ligand–receptor interaction, alcoholism, cAMP signalling and cocaine addiction. For ketamine-associated genes, 78 significantly enriched pathways were found, and 12 pathways were shared between ketamine and CUD (representing 92% of the 13 CUD pathways), indicating substantial pathway-level overlap between ketamine targets and CUD biology.

Gao and colleagues interpret their integrated pipeline as supporting ketamine's potential to improve remission rates in people with CUD. The consistency between the EHR findings and prior small clinical trials — which reported reductions in craving, decreased self-administration and greater abstinence compared with lorazepam or midazolam controls — is emphasised. Unlike those prior trials, which enrolled small, relatively homogeneous samples, the present EHR-based analyses included patients with substantial medical and psychiatric comorbidity and did not find evidence of differential ketamine effects by gender or race, which the authors suggest may indicate broader generalisability. The investigators discuss plausible mechanisms that could underlie an effect of ketamine on cocaine use. Ketamine’s principal pharmacological action is non-competitive antagonism of NMDA receptors, and cocaine has documented effects on NMDA receptor subunits and downstream signalling. Ketamine also influences BDNF, eukaryotic elongation factor 2 and mTOR pathways that modulate synaptic plasticity, processes implicated in both depression and addiction. Pathway enrichment showed substantial overlap between ketamine- and CUD-associated pathways, supporting biological plausibility for ketamine’s impact on addiction-related systems. The authors also note hypotheses that subjective acute effects of ketamine, including mystical-type experiences reported in some studies, might contribute to clinical benefit. Several limitations are acknowledged. Primary among them is the restricted detail available in EHR data: information on drug dosing, duration, route of administration and patient adherence is limited in the available records. The observational, retrospective design means residual confounding and indication bias cannot be excluded despite propensity matching; the analysis was not pre-registered and should therefore be considered exploratory. The authors highlight the need for mechanistic animal studies to delineate ketamine’s biochemical effects on the addicted brain and for larger, randomised clinical trials to definitively assess ketamine’s efficacy and safety for CUD. Finally, the team reflects on the role of AI in this work: while KG-Predict identified ketamine among top candidates, many AI-generated candidates were not considered clinically practical by experts, reinforcing the value of combining computational prediction with expert clinical judgement.