RareSeek R1 Overview

• RareSeek R1 is a specialized large language model designed for clinical reasoning in rare diseases, integrating multi-modal inputs such as EHR narratives, genomic data, and imaging findings.
• It utilizes parameter-efficient fine-tuning via LoRA and chain-of-thought learning to enhance interpretability and boost diagnostic accuracy.
• Graph-augmented retrieval from a Neo4j-based knowledge graph underpins its state-of-the-art performance, improving benchmark outcomes in complex clinical scenarios.

RareSeek R1 is a specialized LLM for clinical reasoning and diagnosis in rare diseases. It combines a domain-tailored, parameter-efficient transformer architecture with multi-modal input capacity, chain-of-thought (CoT) learning, and graph-augmented retrieval to deliver state-of-the-art interpretability and performance on real-world clinical narratives, especially for challenging rare-disease cases (Yang et al., 18 Nov 2025).

1. Model Architecture and Input Representation

RareSeek R1 is based on a 70B-parameter decoder-only transformer architecture (LLaMA-3.3), distilled from a 671B-parameter teacher (DeepSeek-R1) using DeepSeek-R1-Distill-LLaMA-70B as the backbone. Parameter-efficient fine-tuning is implemented via Low-Rank Adaptation (LoRA) modules (rank=8, α=32), which are inserted into each linear layer, freezing base weights and tuning approximately 0.25B new parameters.

Input representations are highly structured:

• EHR Narratives: Tokenized, concatenated by clinical section (chief complaint, history of present illness, family history, physical exam, specialty consults, ancillary testing).
• Phenotypes: Extracted using PhenoTagger, mapped to Human Phenotype Ontology (HPO) identifiers.
• Non-HPO features: Imaging findings, interventions/procedures, functional assessments, laboratory/pathology results, environmental exposures are marked as text spans or encoded as special tokens.
• Genomic Variants: Provided as standardized transcript/codon notation (e.g., NM_000053.4:c.715T>G), linked on-the-fly to the knowledge graph (KG) via GraphRAG.
• Graph-Grounded Retrieval: At each generation step, the model issues Cypher queries to a Neo4j-based rare-disease KG, retrieving and serializing relevant subgraphs as knowledge prompts. Unlike conventional vector retrieval (e.g., $r(q,d) = \frac{q \cdot d}{\|q\|\,\|d\|}$), graph-indexed node retrieval is used for exactness.

2. Training Corpus and Knowledge Graph Integration

Three core resources are used for training:

• RareMed-Corpus: 1.49×10⁵ documents (~500M tokens), consisting of 48,852 de-identified, clinician-confirmed EHRs, 35,722 guidelines and medical texts (ChARD, NORD, Orphanet, OMIM), 30,101 PubMed case reports, and 34,666 phenotype-driven synthetic cases assembled from HPO and Orphanet (Yang et al., 18 Nov 2025).
• RareMed-CoT: 17,477 reasoning chains, initially seeded by 500 expert-annotated EHR cases, expanded via LLM self-generation and expert curation (inter-annotator $\kappa=0.91$).
• RareMed-RAG: Neo4j KG fusing ClinVar, HGMD, HPO, OMIM, Orphanet—totaling $\sim$8,200 diseases, 16,700 phenotypes, 5,200 genes, and 630,000 variants.

Graph nodes encode diseases, phenotypes, genes, and variants, with edges and frequency annotations curated from major rare-disease sources. The Graph Cypher Retriever expands 1–2-hop neighborhoods starting from phenotype/gene/variant identifiers; subgraphs are ranked by information content (IC) for phenotypes and serialized into the prompt at inference.

3. Instruction Tuning and Chain-of-Thought Learning

RareSeek R1 employs a three-phase progressive transfer learning regime:

1. Domain-Specific Instruction Tuning: Models are optimized with AdamW (lr=$10^{-4}$, batch size=4, 3 epochs). Each sample $(x_i, y_i)$ is framed as maximizing:

$$\mathcal{L}_\mathrm{inst} = -\sum_i \log P_\theta (y_i | x_i,\,\texttt{"<instruction>"})$$

using diverse clinical and case-based instructions.

2. Chain-of-Thought Fine-Tuning: For each instance, $(x_i, r_i, y_i)$, the model learns to generate reasoning chains $r_i$ and diagnoses $y_i$:

$$\mathcal{L}_\mathrm{CoT} = -\sum_i \sum_{t=1}^{T_i} \log P_\theta (c_{i,t} | x_i, c_{i,<t})$$

where $P(\mathrm{CT}|x) = \prod_t P(c_t|x, c_{<t})$.

3. GraphRAG-Augmented Reasoning: KG facts are retrieved and prepended to the prompt. Retrieval-augmented scoring combines semantic and graph-based priors:

$$\mathrm{score}(q,d)=\alpha\,\mathrm{sim}(q,d) + (1-\alpha)\mathrm{Prior}_\mathrm{graph}(q,d)$$

with $\mathrm{sim}(q,d)$ as vector similarity and $\mathrm{Prior}_\mathrm{graph}(q,d)$ quantifying presence/weight of KG connections.

Ablation studies show instruction tuning yields a +11.8% (EHR-Internal) and +9.6% (EHR-External) Top-1 improvement, with CoT fine-tuning yielding an additional +9.5% and +7.9% respectively (Yang et al., 18 Nov 2025).

4. Performance and Benchmarking

RareSeek R1 achieves state-of-the-art diagnostic accuracy across multiple public and internal benchmarks:

Benchmark	Top-1 Accuracy	Top-5/Top-10 Accuracy	Notable Features
EHR-Internal (n=4,306)	0.684 ± 0.014	Top-5: 0.839 ± 0.011	In-domain complex hospital records
EHR-External (n=283)	0.719 ± 0.025		Out-of-domain hospital data
RareBench (n=1,197)	0.392 ± 0.025 (vs. GPT-5 0.353)	Top-10: near-identical	Outperforms SOTA on rare-disease cases
MedEHR-Variant (n=147)	+17.0% with GraphRAG (Top-1 0.770)		KG/variant retrieval largest impact
Phenopacket-Store (n=5,213)		Top-10: 0.910 ± 0.007	Multi-modal cohort, post-GraphRAG

Robustness against noisy/overlapping phenotypes is documented—Top-5 accuracies are maintained (0.787–0.840) from sparse to rich key-phenotype cases, while traditional tools (e.g., Exomiser) register Top-5 <0.15 across all tiers. On the 300-hardest complex-phenotype cases, Top-1 is 0.520 (Yang et al., 18 Nov 2025).

Standard diagnostic metrics—precision, recall, $F_1$—are reported as:

$$\mathrm{Precision} = \frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FP}}, \quad \mathrm{Recall} = \frac{\mathrm{TP}}{\mathrm{TP} + \mathrm{FN}}, \quad F_1 = 2 \frac{\mathrm{Precision} \times \mathrm{Recall}}{\mathrm{Precision} + \mathrm{Recall}}$$

5. Human Studies and Interpretability

In 110-case matched studies spanning neurologic and metabolic disorders, RareSeek R1 achieves Top-1 accuracy of 0.473 (mid-level physician parity), with GraphRAG retrieval elevating Top-1 to 0.582 (senior-attending performance range). Clinician-AI collaboration boosts junior, mid-level, and senior Top-1 accuracy by Δ=+0.169, +0.118, and +0.094, respectively.

Auditable decision paths are enabled by explicit CoT reasoning chains, which closely align with established clinical guidelines. Notably, decisive non-phenotypic (non-HPO) evidence constitutes a median of 23.1% (IQR 0.118–0.400) of features in correct diagnoses. The breakdown by category is as follows:

Category	Fraction (%)
Imaging findings	26.5
Interventions	23.0
Functional tests	17.7
Laboratory results	9.1
Molecular tests	8.2
Pathology	7.2
Environmental exposures	5.4

This highlights the multi-modal and evidence-integrated nature of correct rare-disease diagnostic reasoning in practice (Yang et al., 18 Nov 2025).

6. Clinical Impact and Future Directions

RareSeek R1 substantiates a narrative-first, knowledge-integrated paradigm for rare disease diagnosis, significantly shortening the diagnostic odyssey and supporting clinicians with interpretable, knowledge-grounded decision support. Its integration of graph-augmented retrieval and multi-modal reasoning consistently improves both standalone and assistive diagnostic metrics, with auditability and benchmarking rigorously documented.

A plausible implication is that RareSeek R1's methodology—progressive domain tuning, explicit multi-modal CoT, and dynamic use of curated KGs—could serve as a template for future clinical LLMs tasked with similarly heterogeneous, high-stakes domains.

Ongoing and future developments may involve scaling the KG with additional population-genomic and longitudinal evidence, extending synthetic reasoning chains to rarer disease subtypes, and adopting more adaptable prompt and fine-tuning protocols to further boost clinical domain generalization and real-world deployment (Yang et al., 18 Nov 2025).