An Agentic System for Rare Disease Diagnosis with Traceable Reasoning

1 Main

• Rare-disease diagnosis is challenging because disorders are heterogeneous and uncommon, cases are scarce, knowledge changes rapidly, and clinicians need transparent reasoning.The passage notes that approximately 260 to 280 rare genetic diseases are discovered annually and that patients’ diagnostic odysseys average over five years.Source: §1 · pp. 1–2
• DeepRare accepts free-text descriptions, HPO terms and genomic results, then returns ranked candidate diagnoses with reasoning that references verifiable medical evidence.Its three-tier architecture coordinates a central LLM host with specialized phenotype, genotype, normalization and retrieval agents, connected to curated and web-scale medical resources; a self-reflective loop reassesses hypotheses to reduce over-diagnosis and hallucinations.Source: §1 · p. 2
• DeepRare was evaluated on 6,563 cases across eight datasets, 3,134 diseases and 14 specialties, with cases spanning Asia, North America and Europe.Two in-house datasets contributed 330 cases with phenotypes and whole-exome sequencing; diagnoses in this cohort were validated by genetic testing.Source: §1 · p. 2
• DeepRare outperformed 15 methods in HPO evaluations and Exomiser on Xinhua whole-exome cases, while physicians agreed with its reasoning chains in 95.4% of reviewed cases.The passage reports a 57.18% average score and 65.25% at Recall@1, with margins of 23.79% and 18.65% over Reasoning LLM; multi-modal Recall@1 was 69.1% versus Exomiser’s 55.9%. Ten physicians reviewed reasoning across 180 cases.Source: §1 · p. 2
• DeepRare is also deployed as a web-based diagnostic copilot for rare-disease physicians.The paper additionally evaluates different underlying LLMs and analyzes the contribution of individual modules.Source: §1 · p. 2

2.3 HPO-wise Analysis across Dataset

• DeepRare achieved 57.18% average Recall@1 across HPO-wise benchmarks, versus 33.39% for Claude-3.7-Sonnet-thinking, the second-best method.It also scored Recall@1 of 78% and Recall@3 of 85% on RareBench (MME), and 74% and 81%, respectively, on MyGene2; it outperformed all existing methods on all benchmarks.Source: §2.3 · p. 4

2.4 HPO-wise Analysis across Specialties

• DeepRare was superior across almost all specialties, with top-1 accuracy of 60% versus 32% in Endocrine and 49% versus 34% in Digestive cases.Accuracy was highest in Kidneys and Urinary System at 66%, and lower in Lungs and Breathing at 31%.Source: §2.4 · p. 5

2.5 HPO-wise Analysis across Diseases

• DeepRare’s Recall@1 exceeded all baselines for diseases represented by more than 10 cases and was greater than 0.8 for 31.8% of data-sparse diseases.For the latter threshold, DeepSeek-V3 reached 23.5%, DeepSeek-R1 26.6%, and Baichuan-M1 2.5% of tail-end diseases.Source: §2.5 · p. 5

2.6 HPO-wise Comparison against Experts

• DeepRare achieved Recall@5 of 78.5% versus clinicians’ 65.6%, and Recall@1 of 64.4% versus physicians’ 54.6% in a 163-case comparison.Five experienced physicians and DeepRare received the same structured HPO input; physicians could use search and reference materials but not AI tools.Source: §2.6 · pp. 5–6

2.7 Analysis on HPO and Gene Data

• With HPO and genetic data, DeepRare reached Recall@1 of 69.1% on Xinhua and 63.6% on Hunan, exceeding Exomiser’s 55.9% and 58.0%, respectively.Adding genetic information increased Recall@1 from 39.9% to 69.1% on Xinhua and from 33.3% to 63.6% on Hunan.Source: §2.7 · p. 6

2.8 Traceable Reasoning Chain Validation

• Physicians rated DeepRare’s references 95.4% accurate on average across 180 cases sampled from eight datasets.Incorrect references were categorized as hallucinated URLs or irrelevant sources associated with incorrect diagnostic conclusions.Source: §2.8 · p. 6

2.9 Failure Cases Analysis

• Among 200 cases with the correct diagnosis outside the top five, the most common errors were Reasoning Weighing Error (41.0%, 82/200) and Phenotypic Mimic Diagnosis (38.5%, 77/200).Etiologically Associated Diagnosis accounted for 15.5% (31/200), while Reasoning Factual Error and Evidence Linking Error each accounted for 2.5%.Source: §2.9 · pp. 6–7

2.10 Ablation Study

• Agentic design raised average Recall@1 from 26.11% to 54.67% for GPT-4o and from 26.99% to 56.94% for DeepSeek-V3 across five public datasets.Case retrieval, web knowledge and self-reflection contributed complementary strengths; the full system’s dynamic orchestration outperformed partial configurations.Source: §2.10 · p. 7

Discussion

• DeepRare processes chief complaints, genetic data, and detailed phenotypes, generating comprehensive diagnostic reasoning chains to support clinical decision-making.The system was evaluated across diverse datasets, disease categories, and medical centers.Source: §3 · p. 7
• By grounding reasoning in verifiable medical evidence, DeepRare addresses limited rationale in HPO-based tools, genetic-test dependence in some tools, and hallucinations in LLMs.The authors describe this grounding as supporting interpretability and trustworthiness.Source: §3 · p. 8
• DeepRare consistently outperformed existing methods across RareBench, a rare-disease subset of MIMIC-IV-Note, and the Xinhua Hospital dataset, across specialties and input modalities.The study also reports evidence-based reasoning chains with verifiable references, validated through expert clinical assessment.Source: §3 · p. 8
• Evidence-based reasoning with verifiable references could reduce literature-review and case-research time, while broad specialty performance could support non-specialists and resource-limited regions.The authors suggest this potential could help reduce disparities in rare-disease diagnosis.Source: §3 · p. 8

• DeepRare has not fully incorporated potentially valuable data sources, but its MCP-like plugin interface supports future integration of rare-disease knowledge systems and bioinformatics tools.Source: §4 · p. 8
• Its knowledge search agent aggregates phenotypic information; more refined, adaptive retrieval could improve knowledge curation and potentially diagnostic precision.Source: §4 · p. 8
• DeepRare primarily targets patients already aware of rare diseases but without a precise diagnosis; expanding it to non-specialist screening could trigger initial suspicion and broaden clinical impact.A preliminary screening attempt is described in Supplementary Material 13.7, and the MCP-based architecture can integrate new agent servers.Source: §4 · p. 8
• Patient-interaction modules facilitate information gathering, but suitable validation datasets are lacking, so this feature has not yet been experimentally evaluated.Further investigation and iterative improvement are planned as datasets become available.Source: §4 · p. 8
• Future work will expand DeepRare toward rare-disease treatment and prognosis prediction, aiming for a more flexible and comprehensive disease-management ecosystem.Source: §4 · p. 8

6 Author Contributions

• W.Z. and C.W. led computational algorithm design, while Y.F. led clinical design and medical aspects.All authors contributed to study conception and design.Source: §6 · p. 9
• W.Z., C.W. and Y.F. performed data acquisition, analysis and interpretation and contributed to the evaluation framework; W.Z. and C.W. led technical implementation.P.Q., X.Z., Y.Z. and W.X. also contributed to technical implementation; Y.Z., Y.Y., K.S. and W.X. provided technical and infrastructure guidance; Y.F., S.Z., Y.P., Y.Y. and K.S. provided clinical inputs.Source: §6 · p. 9
• All authors contributed to drafting and revising the manuscript, approved its final version and agreed to be accountable for the work’s accuracy and integrity.All listed authors met the ICMJE four criteria for authorship.Source: §6 · p. 9

• The authors declare no competing interests as defined by Nature Portfolio or other interests that might be perceived to influence the paper’s results or discussion.Source: §7 · p. 9

8.1 Problem Formulation

• Patient input may combine free-text reports, structured HPO terms, and a WES-derived VCF, and the system returns ranked disease hypotheses with evidence-grounded rationales.Either phenotype or genotype input may be absent.Source: §8.1 · p. 10

8.2 Main Workflow

• The host orchestrates parallel phenotype and genotype evidence collection, followed by diagnostic decision-making; genotype analysis is skipped when genotype data are unavailable.Phenotype collection standardizes HPO terms, retrieves literature and similar cases, and analyzes phenotypes; genotype collection annotates and ranks variants before host interpretation.Source: §8.2 · p. 10, Information Collection · pp. 10–11

Self-reflective Diagnosis

• The host standardizes tentative disease names to Orphanet or OMIM, retrieves disease-specific knowledge, and self-reflects; if all candidates are ruled out, it repeats collection with greater search depth.It then produces free-text rationales with reference links and verifies URLs, removing invalid links to mitigate hallucination.Source: Self-reflective Diagnosis · p. 12

Phenotype Extractor

• LLM-assisted two-step phenotype extraction improves description accuracy, while BioLORD maps phenotype and disease names to standardized terms and discards matches below cosine similarity 0.8.Phenotype extraction maps patient reports to HPO entities; disease normalization maps predicted names to Orphanet or OMIM items.Source: Phenotype Extractor · p. 13, Disease Normalizer · p. 13

Knowledge Searcher

• The knowledge searcher queries general web and medical sources for current rare-disease information, then uses a lightweight language model to summarize results and filter out non-medical content.Sources include general search engines, academic literature, rare-disease knowledge bases, and medical repositories.Source: Knowledge Searcher · pp. 13–14

Case Searcher

• Case retrieval embeds HPO profiles to select 50 candidates, then reranks them with MedCPT-Cross-Encoder; this two-stage approach outperforms the evaluated alternatives in efficiency and clinical relevance.A lightweight language model subsequently checks retrieved cases for relevance.Source: Case Searcher · p. 14

Phenotype Analyzer

• The phenotype analyzer combines PhenoBrain, PubcaseFinder, and zero-shot LLM inference to generate candidate rare-disease suggestions from structured HPO terms.PhenoBrain and PubcaseFinder each return five suggestions; the LLM also returns five candidates.Source: Phenotype Analyzer · p. 15

Genotype Analyzer

• The genotype analyzer uses Exomiser to annotate and prioritize GRCh37 variants using population frequencies, pathogenicity predictions, variant effects, inheritance modes, and gene-disease associations.Its sources and algorithms include gnomAD and other population databases, PolyPhen-2, SIFT, MutationTaster, OMIM, and HiPhive.Source: Genotype Analyzer · p. 15
• Exomiser candidates retain phenotype, variant, significance, pathogenicity, and disease metadata, while tool results are templated as free text for integration with the central host or other LLM tools.The selected genes are ranked by composite exomiser_score, and curated annotations are passed to the Central Host.Source: Genotype Analyzer · p. 15

8.4 External Data Sources

• DeepRare’s external tier connects medical literature and case resources, including a deduplicated 67,795-case bank and RareBench Task 4’s 1,114-case public subset.The case bank combines published literature, public datasets, and de-identified proprietary cases; RareBench draws from MME, HMS, LIRICAL, and RAMEDIS.Source: §8.4 · p. 16
• Evaluation resources also include Mygene2’s 146 patients across 55 MONDO diseases and a preprocessed DDD cohort of 2,283 cases.Mygene2 cases were limited to confirmed OMIM identifiers and single candidate genes; the DDD cohort was derived from Gene2Phenotype after excluding missing diagnostic results or phenotypes.Source: §8.4 · p. 16

8.5 Clinical Evaluation Dataset Curation

• Clinical curation yielded 9,185 MIMIC-IV-Rare cases and 5,820 Xinhua cases, then assigned recent cases to test sets and historical cases to similar-case libraries.The resulting splits were 1,875 MIMIC test and 7,310 library cases, and 975 Xinhua test and 4,845 library cases; the records retain heterogeneous, potentially noisy real-world phenotypes.Source: §8.5 · pp. 17–18

8.6 Evaluation Datasets Statistics

• Across datasets, rare-disease counts range from 17 to 2,150 and average patient HPO items from 4.0 to 19.4; the benchmark collection covers 2,919 diseases.The authors also calculate average information content, which reflects ontology-term specificity.Source: §8.6 · p. 18

8.7 Baselines

• Baselines span specialized diagnostic tools, general and medical LLMs, and agentic systems, including PhenoBrain, PubcaseFinder, MDAgents [36], and DeepSeek-V3-Search [79].Exomiser is also included as a variant-prioritization baseline combining genomic variants with HPO phenotype terms.Source: §8.7 · pp. 18–19

8.8 Web Application

• The web application guides users through clinical data entry, optional detailed symptom inquiry, HPO mapping with clinician editing, diagnostic analysis, and downloadable reports.Analysis invokes tools and consults literature and case databases; reports can be exported as PDF or Word documents.Source: §8.8 · p. 19

8.9 Ethical Approval and Informed Consent

• The study received Xinhua Hospital ethics approval, obtained written consent before genetic testing, and de-identified all data before research access.Consent covered clinical testing and subsequent research use of de-identified samples, phenotypes, and genomic data.Source: §8.9 · p. 20

9 Data Availability

• The study utilized RareBench, MyGene2, DDD, MIMIC-IV, Xinhua Hospital, and Hunan Hospital data sources.The first four are publicly available; the two hospital datasets are deposited in public repositories under controlled access.Source: §9 · p. 20
• RareBench, MyGene2, DDD, and MIMIC-IV are publicly available through their respective dataset repositories.RareBench is available on Hugging Face; MyGene2 on Harvard Dataverse; DDD through the Decipher website; and MIMIC-IV through PhysioNet.Source: §9 · p. 20
• Xinhua and Hunan Hospital data are deposited at NGDC/CNCB under project PRJCA052720, with variants GVM001237 and GVM001238 and clinical metadata OMIX013512.Genetic and clinical records require a research proposal and home-institute IRB approval submitted to Xinhua Hospital’s Data Access Committee through the NGDC portal.Source: §9 · p. 20

• The figure’s example is Ehlers-Danlos syndrome, under the heading “Output Traceable Reasoning.”Source: §11 · p. 28
• Labeled system components include decision-making, a memory bank, agent orchestration, and LLM-powered functionality.Source: §11 · p. 28
• The diagram also names a disease normalizer, case searcher, case bank, literatures, guidelines, knowledge base, and genetic resources.Source: §11 · p. 28

Human

• 0.954 mean and 1.000 median reference accuracy scores were reported for DeepRare’s traceable reasoning chains.The human-validation figure presents a histogram of reference accuracy scores with a density curve.Source: Figure 4 · p. 31
• Dataset-specific accuracy distributions showed robust performance across eight rare disease datasets.The figure describes these distributions but does not provide dataset-specific values in the caption.Source: Figure 4 · p. 31
• Failure modes were analyzed in a random sample of 200 failed cases from the entire HPO-wise test set.The figure reports a distribution of these failure modes without specifying category counts in the caption.Source: Figure 4 · p. 31
• The ablation study compared five LLMs as central hosts and assessed similar-case retrieval, web knowledge integration, and self-reflection against baseline GPT-4o.The caption describes performance comparisons but supplies no quantitative ablation results.Source: Figure 5 · p. 32

System overview

• DeepRare uses a central host, agent servers, and external medical data sources, with information collection followed by self-reflection on diagnostic results.Inputs can include patient free text, structured HPO IDs, or both.Source: Figure 1 · p. 34

Cohort curation

• MIMIC-IV-Note and Xinhua Hospital curation retained 9,185 and 5,820 cases, respectively, split into 1,875/7,310 and 975/4,845 test/library sets.The multi-center benchmark table describes case distributions, HPO complexity, disease spectrum, provenance, and genetic annotation status.Source: Figure 2 · p. 35, Table 1 · p. 35

Web application workflow

• The web application moves through clinical data entry, systematic inquiry, HPO mapping, diagnostic analysis, and report downloading.Inputs may include demographic and family-history information, manifestations, and optional medical images, lab reports, or VCF files.Source: Figure 3 · p. 36

13.1 Main Workflow of DeepRare System

• DeepRare accepts free text, HPO items, and genetic variants, collects phenotype and genotype information when available, and returns ranked diagnoses with rationales.When genetic data are absent, the workflow uses phenotype-only diagnosis.Source: §13.1 · p. 37

13.3 Reference Link Verification; 13.4 Web Engineering Implementation

• The web implementation uses locally deployed DeepSeek-V3, microservices, caching, synchronized local databases, and parallel processing; reference links are checked for syntax and HTTP reachability.It was deployed and tested in hospitals, with the authors reporting improved diagnostic efficiency and decision-making accuracy under concurrent loads.Source: §13.4 · pp. 38–39

13.5 Ablation Study on Case Bank

• The full DeepRare system outperformed individual configurations, often by >10 points on R@1, with component benefits varying by case-bank coverage.Dynamic orchestration combines complementary components: retrieval yields large gains when similar cases are available, while knowledge tools and reflection help when precedents are sparse or diseases are unseen.Source: §13.5 · p. 39, Table 3 · p. 40, Table 4 · p. 40
• Case retrieval improved R@1 by +60.00 on MME and +34.56 on RAMEDIS, while tool-calling and reflection gained +40.00 on MME over the base LLM.With sparse or no relevant cases, reported improvements were +2.28 on HMS, +8.38 on LIRICAL, and +4.57 on the casebank-disentangled Xinhua subset; similar-manifestation retrieval still helped on unseen diseases.Source: §13.5 · p. 39, Table 3 · p. 40, Table 4 · p. 40

13.6 Phenotype Extractor Comparison

• The final HPO-extraction method reduced errors and increased exact correctness and “Better” physician judgments over DeepSeek-V3 and RAG+DeepSeek-V3.Audited error types included non-canonical terms, associative drift, and lost qualifiers; proposed safeguards include canonical-term normalization, ontology-constrained decoding, and qualifier-preserving summaries.Source: Table 5 · p. 40, §13.6 · p. 40

13.7 Demonstration of Expanding DeepRare with a Screening Module

• 93.94% accuracy and 0.8526 F1 were achieved by the screening extension on 4,825 common-disease/healthy and 1,227 rare-disease cases.The Rare Disease Discriminator is a RoBERTa-Large-based classifier trained on 114,934 curated phenotype descriptions; the authors report that it outperforms a DeepSeek-V3 baseline with very low precision and may reduce false positives.Source: Table 6 · p. 41, §13.7 · p. 41

Prompt 1. Prompt for Baseline LLM Diagnosis; Prompt 2. Prompt for Diagnosis Result Evaluation; Prompt 3. Prompt for discriminate common disease; Prompt 4. Prompt for Rare Disease Type Classification

• The baseline prompt requests five differential diagnoses, the evaluation prompt returns the gold diagnosis rank or “No,” and auxiliary prompts classify rare versus common disease or assign disease categories.The category prompt permits multiple categories and treats slash-separated disease names as synonyms.Source: Prompt 1. Prompt for Baseline LLM… · p. 41, Prompt 2. Prompt for Diagnosis Result… · p. 42, Prompt 3. Prompt for discriminate… · p. 42

Prompt 5. Prompt for Tentative Diagnosis Decision-making; Prompt 6. Prompt for Rare Disease Gene Analysis; Prompt 7. Prompt for Disease Reflection

• The tentative-diagnosis and gene-analysis prompts request ranked rare-disease hypotheses with concise reasoning, source citations, and pathophysiological explanations; gene analysis prioritizes Exomiser findings.A reflection prompt asks the model to assess whether a proposed diagnosis is correct using patient information, similar cases, and disease knowledge.Source: Prompt 5. Prompt for Tentative… · p. 44, Prompt 6. Prompt for Rare Disease Gene… · p. 45, Prompt 7. Prompt for Disease Reflection · p. 46

Prompt 7. Prompt for Disease Reflection

• Disease reflection prompts organize analysis around the patient’s key symptoms, a proposed diagnosis, and supporting evidence from similar cases and medical literature.The inputs are supplied as patient phenotype, similar-case details, and disease knowledge.Source: Prompt 7. Prompt for Disease Reflection · p. 46

Prompt 8. Prompt for Final Diagnosis

• The final-diagnosis prompt asks for the five most likely rare diseases, ranked and explained using patient information, similar cases, primary results, and prior judgments.It requires symptom-linked reasoning, brief pathophysiological mechanisms, and citations to provided references without invented content.Source: Prompt 8. Prompt for Final Diagnosis · p. 47

Prompt 9. Prompt for HPO Extraction

• Phenotype extraction from discharge-note text is limited to the patient’s phenotypes, checked against HPO and returned as HPO identifiers with descriptions.The prompt requests only extracted phenotypes in the specified format.Source: Prompt 9. Prompt for HPO Extraction · p. 48

Prompt 10. Prompt for HPO Extraction Modify

• The modified extraction prompt maps Chinese or English phenotype descriptions to standard English HPO terms while retaining original terms and using “none” for unmatched concepts.It requires valid JSON and excludes phenotypes absent from the input.Source: Prompt 10. Prompt for HPO Extraction… · p. 49

Prompt 11. Prompt for Knowledge Summarization

• The knowledge-summarization prompt asks for a paragraph retaining key messages about phenotype and related disease, or “not a medical-related page” for nonmedical articles.Source: Prompt 11. Prompt for Knowledge… · p. 50

Prompt 12. Prompt for Case Summarization

• The case-comparison prompt asks for a strict Yes-or-No judgment on whether two patients likely have the same disease, based on their phenotype information.Source: Prompt 12. Prompt for Case Summarization · p. 50

Prompt 13. Prompt for Zero-shot LLM Inference

• The zero-shot prompt requests a diverse differential of five rare diseases, ordered from most to least likely, from a clinical case.Source: Prompt 13. Prompt for Zero-shot LLM… · p. 50

Prompt 14. Prompt for Rare Disease Judgment

• The rare-disease judgment prompt requires a strict binary classification, outputting 1 for rare diseases and 0 for common diseases.It permits consulting OMIM or Orphanet and distinguishes rare diseases from rare phenotypes or characteristics.Source: Prompt 14. Prompt for Rare Disease… · p. 51

Case Study: Optic atrophy 1

• DeepRare ranked OPA1-related optic atrophy first for a patient with visual impairment and fundus abnormalities, matching one of the gold diagnoses.Its stop-gained OPA1 variant supports a loss-of-function mechanism; the lower-ranked alternatives had weaker phenotype concordance or required further confirmation.Source: Case Study: Optic atrophy 1 · p. 52

Case Study: HYPOGONADOTROPIC HYPOGONADISM 2

• For testicular atrophy, micropenis, and short stature, DeepRare ranked FGFR1-associated hypogonadotropic hypogonadism first, matching the gold diagnosis.Olfactory assessment can distinguish Kallmann syndrome from normosmic congenital hypogonadotropic hypogonadism; broader 46,XY DSD workup may be warranted if biochemical findings do not support GnRH deficiency.Source: Case Study: HYPOGONADOTROPIC… · p. 53

Case Study: RENAL TUBULAR ACIDOSIS, DISTAL, 4, WITH HEMOLYTIC ANEMIA

• For renal tubular acidosis with growth delay, DeepRare ranked SLC4A1-related dRTA1 first, while the gold diagnosis was dRTA4 with hemolytic anemia.The predicted missense variant was labeled benign/likely benign, so variant review is advised; distinguishing recessive dRTA4 requires checking biallelic status and hemolysis.Source: Case Study: RENAL TUBULAR ACIDOSIS,… · p. 54

Case Study: COFFIN-SIRIS SYNDROME 4

• For developmental delay and microcephaly, DeepRare ranked KMT2C-related Kleefstra syndrome 2 first, although the gold diagnosis was Coffin–Siris syndrome 4.The stop-gained variant and high Exomiser and phenotype scores supported the model’s leading diagnosis; alternatives included CAMK2B- and SMARCA4-related disorders with VUS variants.Source: Case Study: COFFIN-SIRIS SYNDROME 4 · p. 55

Case Study: Intellectual developmental disorder, autosomal dominant 21, CTCF-related neurodevelopmental disorder

• In a CTCF-related neurodevelopmental disorder case, DeepRare ranked Cornelia de Lange syndrome first based on overlapping facial and limb features, but molecular testing was needed to identify the correct diagnosis.Reviewers described the prediction as clinically astute because the distinct genetic cause produced a strikingly similar appearance.Source: Case Study: Intellectual developmental… · pp. 56–57

Case Study: Carbamoyl-phosphate synthetase deficiency

• For neonatal hyperammonemia and cerebral edema, DeepRare ranked OTC deficiency first despite a gold diagnosis of carbamoyl-phosphate synthetase 1 deficiency.The case reasoning linked hypotonia and feeding difficulties to ammonia’s effects on the brain and cited similar neonatal cases.Source: Case Study: Carbamoyl-phosphate… · p. 57

Methylmalonic acidemia (MMA) (Rank #4/5); Propionic acidemia (PA) (Rank #5/5)

• MMA and PA are organic acidemias that can present with hyperammonemia, hypotonia, and feeding difficulties; severe forms can lead to neonatal death, with acid accumulation disrupting metabolism.Source: Methylmalonic acidemia (MMA) (Rank #4/5) · p. 57, Propionic acidemia (PA) (Rank #5/5) · pp. 57–58

Propionic acidemia (PA) (Rank #5/5)

• In the PA case, the golden diagnosis and DeepRare’s top five had high phenotypic overlap, but urine orotic acid needed to distinguish CPS1 deficiency from OTC deficiency was unavailable.Source: Propionic acidemia (PA) (Rank #5/5) · p. 58

Case Study: Focal segmental glomerulosclerosis

• Despite focal segmental glomerulosclerosis as the golden diagnosis, DeepRare ranked CREST first and incorrectly treated FSGS as a symptom, over-weighting secondary inflammatory clues.Its CREST rationale cited cutaneous, vascular, and systemic inflammatory features; the case also notes that urine orotic acid, useful for distinguishing CPS1 from OTC deficiency, was not provided in the separate PA case.Source: Case Study: Focal segmental… · pp. 58–60
• For the FSGS case, the remaining ranked diagnoses were sarcoidosis second, PAN third, FMF fourth, and Alport syndrome fifth.Source: Case Study: Focal segmental… · p. 59

Case Study: Mitochondrial complex IV deficiency, nuclear type 1

• DeepRare ranked methylmalonic acidemia first instead of mitochondrial complex IV deficiency, matching the infantile hypotonia and death but over-weighting secondary dicarboxylic aciduria.The passage identifies distinguishing primary from secondary biochemical abnormalities and balancing laboratory results against the overall presentation as an optimization need.Source: Case Study: Mitochondrial complex IV… · pp. 60–61

Case Study: Hypophosphatasia

• DeepRare ranked primary Sjögren syndrome first rather than hypophosphatasia, favoring ANA positivity and generalized pain over low alkaline phosphatase with severe skeletal findings.The passage describes low alkaline phosphatase in this context as pathognomonic for hypophosphatasia and calls for better weighting of defining features against incidental findings.Source: Case Study: Hypophosphatasia · pp. 61–62

• The template sets genomeAssembly to GRCh37 and requests TSV and HTML output with a configurable output prefix.Source: §13.10 · p. 63
• It configures frequency sources spanning Thousand Genomes, TOPMed, UK10K, ESP, and gnomAD population groups.Source: §13.10 · p. 63
• Variant analysis uses POLYPHEN, MUTATION_TASTER, and SIFT pathogenicity sources with PASS_ONLY mode and keepNonPathogenic set to true.Source: §13.10 · p. 63
• Inheritance weights are 0.1 for autosomal dominant, autosomal-recessive homozygous-alt, and X-recessive homozygous-alt models, and 2.0 for autosomal-recessive and X-recessive compound-heterozygous models.Source: §13.10 · p. 63