RareAgent: Multi-Agent Reasoning for Drug Repurposing

• RareAgent is a self-evolving multi-agent reasoning framework that actively constructs and refines task-specific evidence graphs for drug repurposing in rare diseases.
• It orchestrates four specialized LLM agents to generate, critique, and iterate hypotheses, effectively overcoming sparse-zero graph challenges.
• The framework demonstrates significant performance improvements over baselines while highlighting limitations in scalability, LLM factuality, and noisy data handling.

RareAgent most directly denotes a self-evolving multi-agent reasoning framework for drug repurposing in rare diseases, introduced for the sparse-zero bipartite setting in which neither the queried disease nor the candidate drug has informative drug–disease graph connectivity. In that formulation, the system shifts the task from static score prediction on a biomedical knowledge graph to active construction, adversarial critique, and iterative refinement of task-specific evidence graphs. At the same time, closely related literature uses near-identical names for distinct rare-disease agent systems and, in one case, as a mistaken variant of an unrelated security defense, so the term has acquired a measure of nomenclatural ambiguity across recent arXiv work (Qin et al., 7 Oct 2025, Chen et al., 2024, Wu et al., 14 Jul 2025, Yang et al., 1 Jul 2026, Changjiang et al., 10 Jun 2025).

1. Terminology and referential ambiguity

In the strict sense, RareAgent is the system introduced in "RareAgent: Self-Evolving Reasoning for Drug Repurposing in Rare Diseases" (Qin et al., 7 Oct 2025). However, adjacent papers use similar labels in ways that can easily be conflated.

Term	Referent	arXiv id
RareAgent	Drug repurposing via self-evolving multi-agent reasoning	(Qin et al., 7 Oct 2025)
RareAgents	MDT framework for rare-disease diagnosis and treatment	(Chen et al., 2024)
RDMA	Rare Disease Mining Agents for EHR note mining	(Wu et al., 14 Jul 2025)
"RareAgent" in AgentODE details	Informal guidance label for rare-disease modeling agents	(Yang et al., 1 Jul 2026)
“RareAgent” as mistaken variant	Actually ReAgent, a backdoor defense for LLM agents	(Changjiang et al., 10 Jun 2025)

The distinction matters because the singular RareAgent paper addresses drug repurposing under graph sparsity, whereas RareAgents addresses clinical MDT-style diagnosis and medication recommendation, RDMA addresses rare-disease and phenotype mining from EHR notes, and AgentODE addresses mechanistic ODE discovery and parameter inference from population-level summary statistics. In the ReAgent paper, the authors explicitly state that the string “RareAgent” does not appear in the paper and should be treated as an informal or mistaken variant of ReAgent (Qin et al., 7 Oct 2025, Chen et al., 2024, Wu et al., 14 Jul 2025, Yang et al., 1 Jul 2026, Changjiang et al., 10 Jun 2025).

2. RareAgent in rare-disease drug repurposing

Drug repurposing aims to find new therapeutic indications for existing drugs, and the RareAgent paper frames this as especially important for rare diseases because fewer than 10% of thousands of rare diseases have approved therapies. The work targets the sparse-zero bipartite setting, where standard KG completion and message-passing GNNs have little reliable signal because both endpoints of a candidate drug–disease link are nearly isolated in the graph (Qin et al., 7 Oct 2025).

The underlying biomedical knowledge graph is written as

$G_K = (\mathcal{E}, \mathcal{R}, \mathcal{T}),$

where $\mathcal{E}$ are entities, $\mathcal{R}$ are relation types, and $\mathcal{T}$ are triples. A repurposing query is

$Q = (q, r_{\text{target}}),$

with $q$ a drug or disease and $r_{\text{target}}$ the target relation, such as indication or contraindication. The neighborhood of an entity is

$$\mathcal{N}(e) = \{ v \in \mathcal{E} \mid (e,r,v)\in\mathcal{T} \ \lor \ (v,r,e)\in\mathcal{T} \},$$

and the sparse-zero regime is defined by

$$|\mathcal{N}(q)| \le \epsilon_1 \quad \land \quad |\mathcal{N}(c)| \le \epsilon_2.$$

This formulation is central to the paper’s argument: when both sides of a candidate link are degree-zero or near degree-zero in the drug–disease bipartite subgraph, ordinary message passing, KG embedding, and rule mining are structurally limited. The paper further argues that densifying KGs with similarity or co-mention edges introduces spurious links, dilutes informative edges, and can yield over-smoothing, over-squashing, and poor generalization. RareAgent therefore reframes repurposing from passive pattern recognition on a static graph to active evidence-seeking reasoning (Qin et al., 7 Oct 2025).

The output of a query is not merely a scalar score. Instead, RareAgent defines

$$\mathcal{C}^* = \langle (c_1, G_{E,1}), \dots, (c_k, G_{E,k}) \rangle,$$

where each $\mathcal{E}$0 is a candidate and each $\mathcal{E}$1 is a task-specific evidence graph justifying the proposed link. The learned object is thus a reasoning policy

$\mathcal{E}$2

rather than a static predictor $\mathcal{E}$3 (Qin et al., 7 Oct 2025).

3. Multi-agent architecture and evidence-graph formalism

RareAgent is organized around four LLM-based agents: Principal Investigator (PI), Explorer, Proponent, and Skeptic. The Explorer seeds plausible hypotheses using a KG+GNN retrieval model. The Proponent builds mechanistic and clinical support. The Skeptic searches for counter-evidence, safety concerns, and failure modes. The PI orchestrates the process, scores hypotheses, issues directives, and manages the self-evolution loop (Qin et al., 7 Oct 2025).

The central shared object is the Task-specific Evidence Graph (T-EGraph),

$\mathcal{E}$4

where $\mathcal{E}$5 types nodes as query, hypothesis, or evidence, and $\mathcal{E}$6 types directed edges as entails, supports, or refutes. Each edge also carries a confidence weight $\mathcal{E}$7. Nodes can encode candidate hypotheses, biomedical entities such as genes or pathways, and evidence items such as case reports or mechanistic studies. This makes the graph an auditable record of the reasoning process rather than an implicit latent state (Qin et al., 7 Oct 2025).

The paper’s iterative refinement algorithm begins with Explorer-generated hypotheses and an initial $\mathcal{E}$8. Proponent and Skeptic then propose graph updates, which the PI merges through a graph manager:

$\mathcal{E}$9

Hypothesis nodes are scored by an aggregation operator,

$\mathcal{R}$0

where the arguments include aggregated supporting and refuting weights, mechanistic connectivity, and disjointness of evidence paths. Iteration terminates when rankings stabilize,

$\mathcal{R}$1

or when the round budget is exhausted (Qin et al., 7 Oct 2025).

Conceptually, this architecture makes RareAgent closer to a policy over graph-construction actions than to a conventional link predictor. The Proponent may add chains of the form drug $\mathcal{R}$2 target $\mathcal{R}$3 pathway $\mathcal{R}$4 phenotype $\mathcal{R}$5 disease, whereas the Skeptic may insert contraindications, conflicting studies, or off-target toxicity evidence. The PI can interrupt unproductive exploration, reprioritize safety or mechanistic depth, and reweight the graph-scoring emphasis as the debate evolves (Qin et al., 7 Oct 2025).

4. Self-evolution, textual feedback, and heuristic transfer

A defining feature of RareAgent is its self-evolutionary loop. After each investigation, the PI issues a Credit Assignment Report identifying pivotal, redundant, or misleading steps. This report is used to rewrite each agent’s instructions:

$\mathcal{R}$6

The paper characterizes this as prompt-level policy refinement driven by natural-language critique rather than supervised labels or explicit RL rewards (Qin et al., 7 Oct 2025).

RareAgent also distills successful reasoning paths into a reusable heuristic library. When recurrent patterns appear across high-quality investigations, the PI abstracts them into conditional rules and stores them as

$\mathcal{R}$7

For a new query, relevant heuristics are retrieved and injected into the initial instructions of the Explorer and PI. The intended effect is transfer of “scientific intuition” across related rare-disease investigations without rebuilding all reasoning patterns from scratch (Qin et al., 7 Oct 2025).

The reasoning styles encouraged by this loop are explicitly plural: mechanistic or causal reasoning, clinical and safety reasoning, analogical reasoning across related diseases, literature-based aggregation, and adversarial stress testing. The transparency claim of the paper rests on the fact that the T-EGraph, the debate transcript, and the PI’s scoring rationale remain inspectable. This differs sharply from black-box GNN scores or opaque KG embeddings, because the final candidate ranking is accompanied by an explicit support/refutation structure (Qin et al., 7 Oct 2025).

5. Empirical performance, ablations, and case studies

RareAgent is evaluated on three datasets: a rare-disease drug-repurposing benchmark derived from the 100,000 Genomes Project and clinicaltrials.gov, PrimeKG as the background biomedical KG, and BioHopR for multi-hop biomedical reasoning. After filtering, the 100kGP-based repurposing task contains 230 diseases, 74 drugs, and 1,188 candidate pairs (Qin et al., 7 Oct 2025).

On the rare-disease indication task, the paper reports the following comparison:

Method	AUPRC	AUROC
TxGNN	0.373	0.552
gpt-4o	0.421	0.594
RareAgent	0.438	0.662

RareAgent also attains P@10 = 0.04 and R@10 = 0.27. The paper summarizes this as an 18.1% improvement in indication AUPRC over reasoning baselines and notes that RareAgent surpasses o3-mini (w.Con) despite using a smaller base model. On the contraindication task, all systems have near-zero top-$\mathcal{R}$8 precision and recall, but RareAgent still achieves the best ranking with AUPRC 0.289 versus 0.283 for the best baseline. On BioHopR, RareAgent reports Prec_HOP1 = 33.87, Prec_HOP2 = 17.68, BOTH_COR = 9.74, and BOTH_WR = 56.52, exceeding the reported single-agent LLM baselines on both hop precision metrics (Qin et al., 7 Oct 2025).

Ablation studies show that the paper’s improvements are not attributable to a single component. Full RareAgent achieves AUPRC 0.4381 and AUROC 0.6623. Removing debate yields 0.4067/0.6364; removing the Skeptic yields 0.4314/0.6570; removing PI interrupts yields 0.4145/0.6364; removing textual feedback yields 0.3018/0.5556; and removing heuristic transfer yields 0.3877/0.6518. The largest degradation comes from deleting textual feedback, which the authors interpret as evidence that self-evolution is decisive. Across three self-evolution rounds, AUPRC rises from 0.438 to 0.463 and AUROC from 0.662 to 0.750 (Qin et al., 7 Oct 2025).

The principal qualitative case study is ARVC–Flecainide. RareAgent proposes Flecainide for Arrhythmogenic Right Ventricular Cardiomyopathy, a case in which the graph contains no established drug therapies. The Proponent builds support through sodium-channel blockade and suppression of ventricular ectopy, while the Skeptic highlights pro-arrhythmia and negative inotropic risks, and flags alternatives such as Dronedarone and Digoxin as high-risk or contraindicated. The paper then compares this reasoning path to the later clinical trial NCT03685149, which RareAgent did not access during its run and which reported approximately 72% reduction in ventricular ectopic beats versus placebo with no significant pro-arrhythmic signal. The authors use this case to argue that RareAgent’s evidence-graph and Skeptic-driven pruning can be both predictive and safety-aware (Qin et al., 7 Oct 2025).

6. Other rare-disease systems discussed under similar names

The plural RareAgents names a different framework: a patient-centered, autonomous multi-agent MDT system for rare-disease differential diagnosis and medication recommendation. It simulates a multidisciplinary team of LLM-powered physician agents built on Llama-3.1-8B/70B, uses predefined roles from 41 medical departments, includes dynamic long-term memory and tool use, and is evaluated on RareBench-Public and the newly introduced MIMIC-IV-Ext-Rare dataset. The best reported diagnosis configuration, RareAgents with Llama-3.1-70B, achieves Hit@1 0.5589, Hit@3 0.6867, Hit@10 0.7811, and MR 1.0; on treatment it reports Jaccard 0.4108, F1 0.5563, DDI 0.0796, and #MED 13.17 (Chen et al., 2024).

RDMA, or Rare Disease Mining Agents, is an agentic framework for mining rare diseases and phenotypes from EHR notes. It performs entity extraction, verification, implication reasoning, ontology matching, and dataset refinement; it is designed for local deployment on standard hardware to reduce privacy risk. The paper reports F1 0.65 for phenotype extraction with Mistral 24B in 4-bit quantized form, F1 0.39/0.48/0.59 for rare-disease mention extraction under different label sets, up to 10× reduction in inference cost, and 63% reduction in human re-review workload in refinement, with Cohen’s $\mathcal{R}$9 rising from 0.46 to 0.81 when RDMA is used with human oversight (Wu et al., 14 Jul 2025).

AgentODE is not formally named RareAgent, but its technical exposition explicitly frames itself as guidance for building a “RareAgent” for rare diseases / small cohorts. It jointly discovers ODE structure and infers parameter distributions from population-level summary statistics only, and includes a clinical study on recessive dystrophic epidermolysis bullosa (RDEB) with 231 observations across 46 patients. The paper’s emphasis is mechanistic modeling under privacy constraints rather than diagnosis, repurposing, or EHR coding (Yang et al., 1 Jul 2026).

Finally, the string “RareAgent” appears in discussion of ReAgent, a backdoor defense for LLM agents, but the ReAgent paper states explicitly that the correct name is ReAgent and that “RareAgent” is best interpreted as an informal or mistaken variant. ReAgent operates in a thought–action loop, checks thought–action and instruction–plan consistency, and reports reductions of attack success rate by up to 90% in database operation tasks (Changjiang et al., 10 Jun 2025).

7. Limitations and broader significance

The RareAgent drug-repurposing paper is explicit about several limitations. Its reasoning quality depends on the factuality and reliability of the underlying LLM; hallucination can affect evidence selection and weighting. Distilled heuristics may overfit to particular disease classes or mechanisms. Background resources such as PrimeKG and clinical trials are incomplete and biased toward well-studied entities. Multi-agent debate can struggle in the presence of sparse, noisy, or conflicting evidence. Scalability is also a concern when moving from per-query investigations to much larger disease sets or repeated updates (Qin et al., 7 Oct 2025).

Within those constraints, RareAgent’s broader significance lies in its task reformulation. The paper presents it not as a stronger score function over a fixed graph, but as a framework that converts rare-disease repurposing into active, evidence-seeking reasoning with explicit support, refutation, and entailment structure. This suggests a general pattern for low-resource biomedical inference: when direct relational supervision is weak or absent, performance may depend less on deeper message passing than on the ability to assemble, criticize, and reuse mechanistically plausible evidence paths. In the specific context of rare diseases, that design choice aligns RareAgent with a wider movement toward agentic systems that combine planning, memory, tools, and auditable intermediate reasoning, even though those adjacent systems target different clinical or scientific endpoints (Qin et al., 7 Oct 2025, Chen et al., 2024, Wu et al., 14 Jul 2025, Yang et al., 1 Jul 2026).