RePurpose links plain-language drug repurposing queries to traceable evidence
A new preprint introduces RePurpose, a web-based drug repurposing tool designed to turn a plain-language clinical question into a ranked list of candidate therapies backed by traceable evidence. The platform draws from Open Targets, ClinicalTrials.gov, UniProt, and PubMed, then assembles candidate drugs from structured databases, clinical trial records, or proposals generated by a large language model. The core design goal is simple but clinically relevant: reduce the friction between a bedside hypothesis and the evidence scattered across molecular targets, trial registries, and published literature.
The appeal of that approach is obvious in repurposing work, where the evidence base is often fragmented across several systems that speak different languages. A neurologist considering a remyelinating strategy in multiple sclerosis does not need another black-box ranking engine. What matters is whether the tool can surface the same mechanistic trail a specialist would follow manually, while preserving the source behind each candidate drug.
RePurpose is built around that traceability. Every candidate carries a source that contributes evidence, whether from a structured database, a clinical trial, or a language-model proposal. An adjustable ranking system then orders candidates from more established options to less familiar ones. In practice, that means the platform is not only searching for names, but also exposing how and why a drug entered the list.
Clemastine exposed a sharp split between mechanism-based and disease-based querying
Investigators tested the platform retrospectively using clemastine, a drug that had been considered as a remyelinating agent in multiple sclerosis. The test was revealing. RePurpose retrieved clemastine in every replicate of the mechanism-worded queries, but not in any replicate of the disease-worded queries. When the language model was disabled, no structured query retrieved clemastine at all.
That pattern matters because it shows how query framing can determine whether a repurposing engine reaches the right candidate. A mechanism-based prompt can match the pharmacologic logic of remyelination, target modulation, and pathway alignment. A disease-based prompt, by contrast, may stay too broad and miss a drug whose relevance is mediated through a specific biological pathway rather than the disease name itself.
The clemastine validation also shows where the system is strongest. It appears well suited to tracing evidence around an already formed hypothesis, particularly when the hypothesis is expressed in mechanistic language. In that setting, the platform can connect target directionality to candidate selection and then point back to the supporting source trail.
Ranking mirrored existing evidence, but not later safety concerns
The ranked outputs for clemastine reflected the evidence supporting its candidacy, yet did not capture later evidence of harm in progressive multiple sclerosis. That later safety signal was reachable only through an on-demand literature search. In other words, the tool did not fail because it was indifferent to evidence; it failed in a more specific way, by reflecting the body of evidence available through its current query and ranking architecture rather than independently synthesizing all later contradictory literature.
For clinicians and translational researchers, this creates a practical boundary. A repurposing system can help identify plausible candidates, but it cannot be treated as a substitute for comprehensive literature review when a drug’s history includes evolving efficacy and safety signals. In a field where the same compound may look promising in one disease stage and harmful in another, a ranking engine needs to be read as a map of available evidence, not as a final verdict.
The clemastine example is especially instructive because multiple sclerosis research is often driven by mechanistic hope: remyelination, immune modulation, and axonal preservation can all suggest a path forward long before definitive clinical confirmation arrives. A tool that retrieves a candidate at the mechanistic level but misses later harm can still be useful, but only if users understand exactly what kind of evidence the platform is optimized to surface.
Source traceability is the central clinical feature
The authors frame RePurpose as a source-traceable system rather than a generative answer engine. That distinction is central to its design philosophy. In repurposing, provenance matters as much as novelty. If a drug appears because it has supportive trial data, the user should be able to see the trial trail. If it appears because a large language model proposes it from target relationships, that origin should remain visible as well.
This approach could help separate plausible repurposing leads from unsupported speculation, especially for non-specialists who may not know how to synthesize target data, trial registry entries, and PubMed records in one pass. The platform also allows adjustable ranking, which means users can shift attention between heavily studied candidates and less familiar options. That flexibility may be useful in early hypothesis generation, where established drugs are not always the only interesting ones.
Still, the validation data suggest a narrower claim than broad hypothesis generation. RePurpose appears better at tracing evidence around a hypothesis than inventing one from scratch. That may be the right claim for a clinical research tool, especially in a field where overconfident recommendation engines can create more confusion than clarity.
A repurposing tool that still needs human oversight
The preprint leaves several clinical questions open. It does not establish whether RePurpose can consistently perform across different diseases, different target architectures, or different therapeutic classes. It also does not show how the platform would handle conflicts between mechanistic plausibility and late safety data beyond the clemastine example. Those are not small gaps for a repurposing engine; they define the margin between a helpful search aid and a misleading one.
What the validation does establish is more focused: mechanism-worded prompts can outperform disease-worded prompts when the goal is to recover a known candidate with a specific biological rationale. For multiple sclerosis, that may be exactly the kind of search behavior a translational team wants when exploring remyelinating strategies. For drug safety, however, the platform’s dependence on the query structure and evidence inputs means final review still has to come from a clinician or researcher who can weigh new literature against older promise.
The tool is freely accessible online, and the code is available through the project repository. Those access points make the platform easy to inspect, but they also invite scrutiny from users who will want to know how ranking, source selection, and language-model proposals interact before the system is trusted in real-world repurposing work.