RePurpose: An LLM-Assisted, Source-Traceable Drug Repurposing Tool with Retrospective Validation on Clemastine in Multiple Sclerosis
Abstract
Summary Thus far, evidence for drug repurposing is contained in target-, drug-, trial-, and literature-based data sources that are challenging to synthesize for non-specialists. Here, we introduce RePurpose, a web-based tool that takes as input a topic written in plain language and returns candidate drugs built from data in Open Targets, ClinicalTrials.gov, UniProt, and PubMed. For queries in the form of a goal, RePurpose filters candidate drugs by whether they modulate all targets implicated in the goal in the corresponding direction. Every candidate has a source that contributes evidence: a structured database, a clinical trial, or a proposal from a large language model. An adjustable ranking adjusts the order of candidates from well-studied to less well-known candidate drugs. We retrospectively tested RePurpose with the trial drug clemastine, which was being considered as a remyelinating agent in multiple sclerosis (MS). RePurpose retrieved clemastine in each replicate of the mechanism-worded queries, but in no replicates of the disease-worded queries, and no structured query retrieved it without the language model enabled. Its ranked outputs reflected the evidence supporting clemastine but not later evidence of harm in progressive MS, which was reachable only through an on-demand literature search. Our results imply that RePurpose may be better at tracing evidence around a hypothesis than generating new hypotheses. Availability and Implementation RePurpose Pharma is freely accessible at https://www.repurposepharma.com/. For code and more information on the application, see: https://github.com/nayagamnitin-hub/RePurpose-Pharma.
02 /Specification & Classification
Click any tag to browse matching papers- Citations
- 0
- Organism
- Research Focus
- Approach
In-depth summary
Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (16 Oct 2026) · CC BY 4.0
Large-language-model-assisted repurposing web tool recovered clemastine for multiple sclerosis remyelination only through mechanism-worded, model-dependent queries; its rankings missed later evidence of harm in progressive disease.
Evidence relevant to repurposing sits in separate target, drug, trial, and literature resources with different identifiers, so non-specialists find it hard to combine by hand. RePurpose combines free-text input, direction-of-effect filtering, per-candidate source labels, and an ExEmPLAR-inspired tunable ranking, then tests whether it rediscovers a trial-tested hypothesis and its later harm signal.
Key results
ShownRecovery of clemastine depended entirely on the language model. It appeared only when queries described a remyelination mechanism (goal mode). It never appeared for queries that named the disease, and never with the model switched off.
Goal mode: clemastine present in 5/5 cache-disabled runs of each of two queries. Disease mode ('multiple sclerosis', 'demyelinating disease'): absent in every run. Model off: goal queries returned no candidates and disease queries lacked clemastine. It entered via trial records (NCT05338450, NCT05131828) or database entries, never as a model proposal. · 3.3 Recovery of clemastine; 3.6 Repeatability and contribution of the…; 4.1 Interpretation of the case study
ShownDisease mode could not reach clemastine for two reasons. Open Targets' top multiple sclerosis targets reflect immune pathology rather than repair biology. The trial search returned only a few large trials of disease-modifying therapies per condition.
The top-30 targets, all above the 0.1 association cutoff, were almost exclusively immune or inflammatory (e.g., NR3C1, MS4A1, sphingosine-1-phosphate receptors, integrins). CHRM1, whose antagonism underlies clemastine's remyelinating effect, and CHRM3 were absent. Small remyelination-specific clemastine trials were excluded. · 3.3 Recovery of clemastine; 4.1 Interpretation of the case study
ShownThe ranked scores and the conversational assistant reflected the evidence supporting clemastine. Later harm in progressive MS was absent from scores and surfaced only through an on-demand PubMed search, which did not return the rabbit progenitor-depletion study.
Clemastine got a safety score of 0.70, even though TRAP-MS closed its clemastine arm after 3 of the 9 progressive-MS patients met disability stopping criteria. The TRAP-MS paper topped the View studies results; the rabbit study did not appear. The assistant cited only the original screen and ReBUILD. · 3.4 Contrary evidence; 4.1 Interpretation of the case study; 4.2 Limitations
ShownClemastine's rank and source class varied across repeat runs, and its rank differed widely between the two goal-mode queries, so a rank from a single run is a weak signal of plausibility.
Single runs: 12/80 (confidence 0.61) for 'promote remyelination in multiple sclerosis' vs 76/80 (confidence 0.54) for the oligodendrocyte precursor query. Across 5 runs each: ranks 15th to 55th and 6th to 33rd of 80, with the source switching between trial (phase 3) and database (phase 4) entries. · 3.3 Recovery of clemastine; 3.6 Repeatability and contribution of the…; 4.1 Interpretation of the case study
ShownThe novelty slider moved clemastine non-monotonically. It lifted clemastine above approved drugs at moderate novelty weights but pushed it below early-stage and model-proposed candidates at the extreme.
Clemastine entered with source weight 0.6, phase 3 and novelty 0.425. Re-ranking the 'promote remyelination' run at α = 0, 0.25, 0.5, 0.75, 1.0 moved it from 15th of 80 to 9th, back to 12th, then down to 21st. · 3.5 Effect of tunable ranking; 4.1 Interpretation of the case study
Methods, data and reproducibility
Claude Sonnet 5 routes a free-text query to disease mode (Open Targets disease-target associations) or goal mode (model-proposed targets with direction and weight); candidates from Open Targets, ClinicalTrials.gov, and model proposals pass direction or relevance checks, are deduplicated, scored heuristically then rescored by the model, and reordered by a novelty slider.
- Open Targets Platform disease-target associations (resolved via EFO; targets scoring below 0.1 dropped) and drug mechanism, clinical-stage and safety records · n = top 30 targets retained per disease query
- ClinicalTrials.gov drug and biologic interventions from trials whose listed conditions match the query; clemastine linked to two trial entries · human (clinical trial registrations) · NCT05338450; NCT05131828
- UniProt reviewed Swiss-Prot entries giving target protein names and function text · human (limited to reviewed human entries)
- PubMed records retrieved through NCBI E-utilities for the View studies and Ask AI features
- Retrospective case-study runs: two disease-mode and two goal-mode queries plus two assistant questions; repeated with cache disabled and with the model off · n = 5 runs per structured query (10 goal-mode runs)
- Published clemastine evidence used as the reference: ReBUILD and CCMR-Two phase 2 trials, TRAP-MS progressive-MS arm, rabbit chronic demyelination model · human; rabbit · n = ReBUILD 50 patients; CCMR-Two 67 patients; TRAP-MS arm 9 patients
- Methods & tools
- RePurpose (RePurpose Pharma web tool), Claude Sonnet 5 (anthropic/claude-sonnet-5) via OpenRouter, Open Targets Platform, Experimental Factor Ontology (EFO), ClinicalTrials.gov, UniProt / Swiss-Prot, PubMed via NCBI E-utilities, n8n workflow (conversational assistant), Claude Code (code development and case-study analysis)
- Data & accessions
- http://clinicaltrials.gov/
- Software & environment
- Claude Sonnet 5 (anthropic/claude-sonnet-5) via OpenRouter; RePurpose commit 81f67e4 (2026-10-10); runs on 2026-10-13 with result cache disabled, candidate-proposal temperature 0.2, default slider α = 0.
- Experimental conditions
- Each cache-disabled structured query took 250 to 430 seconds; hardware not stated.
- Reproducibility gaps
- Outputs are stochastic (temperature 0.2 sampling, model rescoring of roughly the top 45 candidates). The tool depends on an external LLM API (OpenRouter) and on live Open Targets/ClinicalTrials.gov contents. Prompts and reports are only in Supplementary File S1.
Limitations (6)
- Validation rests on a single retrospective drug-indication pair. There is no benchmark of approved and failed repurposings (e.g., repoDB) and no comparison with existing repurposing tools, and results depend on the state of the databases at run time. · stated by the authors
- Every route that recovered clemastine passed through a model that was likely trained on the clemastine literature. The goal queries were also written with the expected mechanism in mind, so recovery may reflect recall rather than hypothesis generation. · stated by the authors
- Single-step disease-to-target retrieval cannot reach drugs that act outside a disease's association profile (e.g., via CHRM1). The trial search keeps only a few top trials per condition, so it misses small, mechanism-specific studies. · stated by the authors
- Scoring weights were chosen by hand without outcome validation. Model rescoring of roughly the first 45 candidates puts two scoring regimes on different scales in one list, and 5 runs are too few to estimate rank variability. · stated by the authors
- Evidence against a candidate is ignored: trial records carry only existence and phase, and safety scores are not specific to indication or disease stage. The novelty measure penalizes late-stage drugs, and the assistant was tested with only two questions. · stated by the authors
- Clemastine's default-order rank for the 'promote remyelination' query is 12/80 in Section 3.3 but 15th of 80 at α = 0 in Section 3.5. The case-study analyses were produced with Claude Code, so Supplementary File S1 should be checked directly. · noted in this summary
Who should read:Read this if you build or evaluate LLM-assisted drug repurposing tools and want a candid case study on recall versus discovery, source labeling, and missed evidence of harm.
Reproducibility ledger
Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.
Sign in to file a reproducibility report.
Related papers
- DReAmocracy: A Method to Capitalize on Prior Drug Discovery Efforts to Highlight Candidate Drugs for Repurposing
Kyriaki Savva, Margarita Zachariou, … George M. Spyrou·International Journal of Molecular SciencesPeer-reviewed·May 13, 2024
- Systematic AI-Driven Drug Repurposing via Clinical Trial Data Mining: A Framework and Six Cross-Therapeutic Case Studies.
Vrinda Gote·bioRxiv·Jun 14, 2026
- Drug repositoning or target repositioning: a structural perspective of drug-target-indication relationship for available repurposed drugs.
Daniele Parisi, Melissa F. Adasme, … Michael Schroeder·Computational and Structural Biotechnology JournalPeer-reviewed·Jan 1, 2020
- From General-Purpose to Disease-Specific Features: Aligning LLM Embeddings on a Disease-Specific Biomedical Knowledge Graph for Drug Repurposing
S. C. Pandey, Muhammed Talo, … Serdar Bozdag·bioRxiv·Mar 10, 2026
- An Integrated Knowledge Graph and Network Medicine Pipeline for Drug Repurposing: Benchmarking Across Human Diseases and Application to Amyotrophic Lateral Sclerosis
Anqi Jiang, Jiajing Hu, … Alfredo Iacoangeli·bioRxiv·Jul 8, 2026
- Drug repurposing through joint learning on knowledge graphs and literature
Mona Alshahrani, Robert Hoehndorf·bioRxiv·Aug 6, 2018