Genosolver: large language models that read clinical notes to diagnose rare diseases — causal gene ranked first in 72% of solved cases, but only 1.7% more diagnoses on unsolved ones

A team from the Center for Human Genetics at RWTH Aachen University builds Genosolver, a chain of large language models that reads raw clinical reports, extracts patient features, then re-ranks genetic variants to propose a rare-disease diagnosis. On 233 already-solved cases, the causal gene is ranked first in 72% and among the top ten in 94%, ahead of the reference tool Exomiser. But applied to 1,875 cases that had remained undiagnosed, the tool resolves only 1.7% more, the comparator receives less information than the model, and only one of the large models tested beats the reference.

The context

A rare disease affects few people taken individually, but there are thousands of them and, together, they concern tens of millions of patients. Many have a genetic cause. The diagnostic journey — the "diagnostic odyssey" — often lasts years, and despite exome sequencing (WES, reading the coding regions of the genome) or whole-genome sequencing (WGS), a majority of patients remain without an answer after a state-of-the-art work-up.

The bottleneck is not only genomic, it is phenotypic. To match a patient to a gene, clinical signs are summarised into standardised terms of the Human Phenotype Ontology (HPO, a controlled vocabulary that gives each symptom a unique identifier). But HPO captures poorly the age of onset, the evolution over time, the detail of a symptom, family history, or the absence of a sign — all information that lives in the free text of reports, not in a list of codes. The promise of large language models (LLMs, models trained to produce text, able to read and summarise documents) and reasoning models (LRMs, LLMs trained to lay out reasoning steps before answering) is precisely to read this raw text. This preprint's question: does that reading really improve diagnosis, and by how much under real conditions?

The method

Genosolver is a chain of three modules, with no retraining of the models, enriched by a RAG approach (retrieval-augmented generation: the model first fetches relevant passages from a document base, then reasons from them rather than from memory alone). The first module extracts phenotypic features from unstructured records. The second builds a list of candidate diseases by filtering a knowledge base (GeneReviews, UniProt, HPO) via vector search. The third matches these hypotheses to the patient's variant list and re-ranks them, producing a reasoning summary.

The knowledge bases queried are massive: 102,899 gene-phenotype lines derived from the 950 chapters of GeneReviews, 12,412 gene-disease associations from UniProt covering 12,260 genes. The clinical text, in German, is translated to English by a dedicated model. For the final prioritisation, the authors compare several locally runnable LLMs — Llama-3.3-70B, MedGo 32B, DeepSeek-R1-Distill-70B, GPT OSS 120B — plus GPT-5.1 via an EU-resident service, a choice driven by health-data confidentiality (everything runs locally or within the EU).

The main evaluation is on 233 cases already genetically solved at Aachen (neurodevelopmental, neuromuscular, organ-specific and haematological conditions; median age 25.5 years). The comparator is Exomiser, an established variant-prioritisation tool, run with its standard settings and expert-assigned HPO terms. The metric is the hit rate at k (recall@k): the proportion of cases where the correct gene is among the top k of the list, for k = 1, 3, 5, 10. A second cohort serves as a full-scale test: 1,875 previously unsolved cases, reanalysed two to three years after their sequencing.

The results

On the 233 solved cases, the best configuration — Genosolver fed the raw notes and coupled with GPT OSS 120B — ranks the causal gene first in 72% of cases (recall@1 = 0.72), among the top three in 83%, among the top ten in 94% (recall@10 = 0.94). Exomiser, the second best, reaches 0.63 / 0.78 / 0.85 at the same ranks: the gap is about nine points at recall@10. Notably, this result rests on a single model: the three other LLMs tested do not beat the reference. When raw notes are replaced by HPO terms alone, Genosolver's advantage melts away, confirming that the gain comes from reading free text, not from the architecture.

Reading the text brings real information: 522 extracted phenotypic elements could not be attached to any HPO term, including 342 traits simply missed by the experts, 67 temporal items (onset age, evolution), 8 negative traits (the explicit absence of a sign) and 6 family-history elements. The reasoning model exploits these cues: it uses age in 43 cases (100% appropriately), family history in 48 cases (47 correct), but sex in only 9 cases, 3 of them wrongly — a trace of hallucination.

Clinical translation, and this is the number that matters. On the 1,875 truly unsolved cases, Genosolver produces 31 new diagnoses, an added yield of 1.7% — an order of magnitude consistent with other reanalysis pipelines. And of those 31, 21 came down to a simple adjustment of variant-filtering rules, 6 to a reclassification in ClinVar in the meantime, 2 to a change in gene annotation. In other words, most of the "real-world" gain comes not from new diagnostic reasoning but from the evolution of annotations and thresholds. There is a long way from the showcase 72% to the field's 1.7%.

What is good

Capturing the phenotype beyond the code list. The most solid contribution is measured, not merely asserted: 522 clinical elements absent from HPO, finely categorised (missed traits, temporal dynamics, negative signs, family history). This is the LLMs' real contribution here — reading what standardisation discards — and the fact that the advantage evaporates when reverting to HPO terms alone demonstrates it by contrast.

An established comparator and a concern for confidentiality. Many LLM papers compare themselves to themselves. Here the reference is Exomiser, a tool genuinely used in clinical genetics, tested across several fusion configurations. And the whole thing is designed to run locally or with EU residency, a concrete constraint rarely respected when handling patient records.

Honesty about the real-world yield. The authors do not hide the 1.7% behind the 72%. They report the reanalysis of unsolved cases, break down the origin of the 31 diagnoses, and situate their figure against existing pipelines. This distinction between re-ranking a list and discovering a diagnosis is exactly what most announcements in the field lack.

What is less good

An evaluation by proxy exposed to annotation leakage. This is the central limitation, and it involves both data leakage and a misleading metric. The 233 cases are already solved; the bases queried by the RAG (GeneReviews, OMIM, ClinVar) already contain the gene-disease association for these cases; and the variant list submitted to the model is pre-filtered on pathogenic annotations (non-benign ClinVar, CADD above 20). The correct gene is therefore almost guaranteed to be present in the input list: recall@k measures an ability to reorder a short, already-enriched list, not to diagnose from scratch. The only test that escapes this bias — the unsolved cases — yields 1.7%.

A disadvantaged comparator. The "+9%" over Exomiser is not obtained on a level playing field: Exomiser receives only the experts' HPO terms, whereas Genosolver receives the full rich clinical notes. Two tools fed different information are compared, and the gap is then attributed to the second. With equal input (HPO alone), the advantage disappears — the authors show it themselves.

A single site, a single model, closed reproducibility. All validation is single-centre (Aachen, translated German text): no external cohort, no population diversity, unknown generalisability. Only one of the four LLMs tested beats the reference, making the conclusion fragile with respect to model choice. No confidence interval or statistical test accompanies the gaps. Finally, the code is only promised "upon publication," patient data are not shareable, and this is a preprint not peer-reviewed, which the authors themselves flag as not to be used to guide clinical practice.

What it changes

For the research community, the useful message is not "LLMs diagnose rare diseases" but "LLMs can extract from free clinical text phenotypic elements that HPO lets slip." That is a preprocessing contribution, real and reusable. The preprint also offers a methodological lesson: evaluating such a tool on already-solved cases, with pre-annotated variant lists, mechanically overestimates its value; the only credible judge is a truly unsolved cohort, where the figure falls back to 1.7%.

For clinicians, nothing is deployable today: a semi-automated research tool, single-centre, unapproved, without prospective validation, and explicitly not intended for practice. The idea of an assistant that re-reads reports to lose nothing of the phenotype is credible in the medium term, in support of the geneticist, not in their place. For patients and the general public, the study illustrates the gap between a headline and a reality: a spectacular "72%" describes a re-ranking of lists where the answer is already present, whereas the real-life gain — solving open cases — is 1.7%. Which takes nothing away from the value of reading records better: for a family lost in the odyssey for ten years, one extra diagnosis in fifty-nine is not nothing.

To go further

The preprint is available on medRxiv (10.64898/2026.06.04.26354845). On large language models set against diagnostic reasoning, see our decryptage of an LLM that actively seeks information before concluding; on an AI diagnostic system validated across multiple centres, that of a tool detecting 22 pediatric respiratory pathogens; and on the gap between headline performance and real-world validation, that of a real-world validation of an adherence model.

Editorial transparency: French version written and signed by the Tatakoto editorial team based on a reading of the preprint. English, Spanish and Chinese translations produced with AI assistance and reviewed.