Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 127 of the 127 references it cites were found in a registry or an index; 32 of the 49 sources its claims cite were checked for support.

Few shot learning for phenotype-driven diagnosis of patients with rare genetic diseases

Open access
Alsentzer, Emily, Li, Michelle M., Kobren, Shilpa N., Noori, Ayush, Undiagnosed Diseases Network, Kohane, Isaac S., Zitnik, Marinka · 2025 · npj Digital Medicine

Published version of a preprint

Abstract

There are over 7000 rare diseases, some affecting 3500 or fewer patients in the United States. Due to clinicians' limited experience with such diseases and the heterogeneity of clinical presentations, ~70% of individuals seeking a diagnosis remain undiagnosed. Deep learning has demonstrated success in aiding the diagnosis of common diseases. However, existing approaches require labeled datasets with thousands of diagnosed patients per disease. We present SHEPHERD, a few-shot learning approach for multi-faceted rare disease diagnosis. SHEPHERD performs deep learning over a knowledge graph enriched with rare disease information and is trained on a dataset of simulated rare disease patients. We demonstrate SHEPHERD's effectiveness across diverse diagnostic tasks, performing causal gene discovery, retrieving "patients-like-me", and characterizing novel disease presentations, using real-world cohorts from the Undiagnosed Diseases Network (N = 465), MyGene2 (N = 146), and the Deciphering Developmental Disorders study (N = 1431). SHEPHERD demonstrates the potential of knowledge-grounded deep learning to accelerate rare disease diagnosis.

Overview

CoChat Summary

TL;DR

SHEPHERD is a few-shot geometric deep learning method that uses a rare disease knowledge graph and simulated patients to perform phenotype-driven causal gene discovery, “patients-like-me” retrieval, and novel disease characterization. Evaluated on real cohorts including 465 Undiagnosed Diseases Network patients, it often ranks causal genes highly and generalizes across sites, diseases, and atypical presentations.

Key findings

  • On UDN EXPERT-CURATED gene lists, SHEPHERD ranked the causal gene first in 40% of patients, with average recall 0.69 at k = 3 and 0.85 at k = 5; on longer VARIANT-FILTERED lists, average recall was 0.21, 0.38, and 0.48 at k = 1, 5, and 10.
  • For hard-to-diagnose patients whose phenotype terms were more than two hops from the causal gene in the knowledge graph, SHEPHERD identified the correct causal gene among its top five predictions for 77.8% of patients.
  • In hard scenarios, SHEPHERD achieved win rates up to 82% and 83% for patients whose causal genes lacked known phenotype or disease associations, respectively, on EXPERT-CURATED lists, and up to 86% on VARIANT-FILTERED lists for novel disease genes.
  • SHEPHERD’s patient embeddings captured disease similarity (adjusted mutual information = 0.304; p < 0.01) and enabled “patients-like-me” retrieval, ranking a MyGene2 patient with the same causal gene first for 11.5% of UDN patients and achieving recall 0.31, 0.43, 0.49, and 0.53 at k = 5, 10, 25, and 50.

Methods

  • SHEPHERD uses a rare disease knowledge graph with 105,220 nodes and 1,095,469 edges, built from sources including PrimeKG, HPO, MONDO, Orphanet, DisGeNET, GO, and Reactome.
  • It is pretrained via self-supervised link prediction on the graph, then fine-tuned for diagnosis tasks using a graph attention network, patient phenotype subgraphs, and disease-stratified training on 42,624 simulated patients representing 2,132 rare diseases.
  • Evaluation used real cohorts: UDN (N = 465), MyGene2 (N = 146), and DDD (N = 1431), with causal gene discovery compared against domain-specific tools, language models, traditional machine learning, network science baselines, and random ranking.

Limitations

  • The paper states that continually updating the knowledge graph could improve performance, that still-undiagnosed UDN patients may be more challenging than diagnosed patients tested, and that datasets likely overrepresent individuals of European descent, limiting generalizability.
  • The text also makes evident that evaluation is retrospective on already-diagnosed patients and does not assess prospective clinical workflow impact; the authors note prospective evaluation as an important next step.
  • The paper states that SHEPHERD relies on structured phenotype terms and a knowledge graph, and suggests extensions such as incorporating variant-level information, negative phenotypes, temporal modeling, and multimodal data.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s41746-025-01749-1

Cites

127 of 127 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 Overview of SHEPHERD in the rare disease diagnosis pipeline. a After years of failed diagnostic attempts, once a patient is accepted to the UDN, they receive a thorough clinical workup and genetic sequencing, and their case is analyzed in an iterative process to identify the candidate genes likely to explain the patient’s symptoms. SHEPHERD can be used throughout the diagnostic process: after the clinical workup to find similar patients, after the sequencing analysis to identify strong candidate genes, and after the case review to further prioritize candidate genes, characterize the patient’s disease, and/or validate candidate genes by finding phenotype- and genotype-matched patients. b SHEPHERD takes in as input the patient’s set of phenotype terms and leverages an external rare disease knowledge graph to perform multi-faceted rare disease diagnosis. SHEPHERD can optionally consider a list of candidate genes (either variant-filtered or expert-curated) or external patient cohort(s), depending on the prediction task of interest (e.g., causal gene discovery, patients-like-me identification). For simplicity, the knowledge graph is depicted using three shapes: circles as genes, squares as phenotypes, and pentagons as diseases; refer to Methods for all node types. c Number of HPO phenotype terms and candidate genes in each of the two candidate gene lists across patients in our UDN cohort. d Overlap of phenotype terms, genes, and diseases across patients. Most phenotype terms, genes, and diseases are found in only a single UDN patient. e–h Number of patients in each e UDN clinical site, f age category, g primary presenting symptom, and h evaluation year. Figure adapted from images created in BioRender129–131.
Fig. 1 Overview of SHEPHERD in the rare disease diagnosis pipeline. a After years of failed diagnostic attempts, once a patient is accepted to the UDN, they receive a thorough clinical workup and genetic sequencing, and their case is analyzed in an iterative process to identify the candidate genes likely to explain the patient’s symptoms. SHEPHERD can be used throughout the diagnostic process: after the clinical workup to find similar patients, after the sequencing analysis to identify strong candidate genes, and after the case review to further prioritize candidate genes, characterize the patient’s disease, and/or validate candidate genes by finding phenotype- and genotype-matched patients. b SHEPHERD takes in as input the patient’s set of phenotype terms and leverages an external rare disease knowledge graph to perform multi-faceted rare disease diagnosis. SHEPHERD can optionally consider a list of candidate genes (either variant-filtered or expert-curated) or external patient cohort(s), depending on the prediction task of interest (e.g., causal gene discovery, patients-like-me identification). For simplicity, the knowledge graph is depicted using three shapes: circles as genes, squares as phenotypes, and pentagons as diseases; refer to Methods for all node types. c Number of HPO phenotype terms and candidate genes in each of the two candidate gene lists across patients in our UDN cohort. d Overlap of phenotype terms, genes, and diseases across patients. Most phenotype terms, genes, and diseases are found in only a single UDN patient. e–h Number of patients in each e UDN clinical site, f age category, g primary presenting symptom, and h evaluation year. Figure adapted from images created in BioRender129–131.CC BY-NC-ND · PubMed Central
Fig. 2 SHEPHERD architecture, training, and generalizability. a, b SHEPHERD is trained in a two-step process. a First, the model is pretrained to embed the biomedical knowledge in the knowledge graph. b Then, the pretrained model is applied to the task of rare disease diagnosis. Patient information is overlaid on the knowledge graph, and SHEPHERD generates an embedding for the patient phenotype terms and each candidate gene, disease, or patient. The model is trained via a loss function that encourages patient embeddings to be close to the embeddings of their causal gene or disease or other patients with the same causal gene or disease. c SHEPHERD is trained on a large cohort of simulated patients (pink). It can be further trained on real-world patients (blue) and then evaluated on an independent cohort of real-world patients (green). Alternatively, SHEPHERD can directly be evaluated on real-world patients (green) without any additional training. d We leverage real patient data derived from three distinct cohorts: the Undiagnosed Diseases Network (UDN; N = 465), MyGene2 (N = 146), and Deciphering Developmental Disorders study (DDD; N = 1431). For simplicity, the KG is depicted using three shapes: circles as genes, squares as phenotypes, and pentagons as diseases; refer to Methods for all node types.
Fig. 2 SHEPHERD architecture, training, and generalizability. a, b SHEPHERD is trained in a two-step process. a First, the model is pretrained to embed the biomedical knowledge in the knowledge graph. b Then, the pretrained model is applied to the task of rare disease diagnosis. Patient information is overlaid on the knowledge graph, and SHEPHERD generates an embedding for the patient phenotype terms and each candidate gene, disease, or patient. The model is trained via a loss function that encourages patient embeddings to be close to the embeddings of their causal gene or disease or other patients with the same causal gene or disease. c SHEPHERD is trained on a large cohort of simulated patients (pink). It can be further trained on real-world patients (blue) and then evaluated on an independent cohort of real-world patients (green). Alternatively, SHEPHERD can directly be evaluated on real-world patients (green) without any additional training. d We leverage real patient data derived from three distinct cohorts: the Undiagnosed Diseases Network (UDN; N = 465), MyGene2 (N = 146), and Deciphering Developmental Disorders study (DDD; N = 1431). For simplicity, the KG is depicted using three shapes: circles as genes, squares as phenotypes, and pentagons as diseases; refer to Methods for all node types.CC BY-NC-ND · PubMed Central
Fig. 3 SHEPHERD performs generalizable causal gene discovery. a Performance of SHEPHERD, four domain-specific approaches, five language model, traditional machine learning, and network science baselines, and a random baseline. The performance metric is average recall at k for k = 1, 3, and 5. Error bars denote standard deviation over models trained with five random seeds. b, c Performance of SHEPHERD in ranking causal genes stratified by b clinical sites and c primary presenting symptoms. Each boxplot shows the median and interquartile range of the rank of the causal gene. Whiskers extend to ±1.5 × IQR. d Performance of SHEPHERD, six domain-specific approaches, five large language model, traditional machine learning, network science baselines, and a random baseline. The performance metric is average recall at k for k = 1, 5, 10, 25, and 50. Error bars denote standard deviation over models trained with five random seeds. e, f Performance of SHEPHERD against domain-specific algorithms in four extremely hard-to-diagnose scenarios on e EXPERT-CURATED and f VARIANT-FILTERED gene lists. Shown is the win rate, the proportion of patients where SHEPHERD performs the same as or better than the benchmark algorithms.
Fig. 3 SHEPHERD performs generalizable causal gene discovery. a Performance of SHEPHERD, four domain-specific approaches, five language model, traditional machine learning, and network science baselines, and a random baseline. The performance metric is average recall at k for k = 1, 3, and 5. Error bars denote standard deviation over models trained with five random seeds. b, c Performance of SHEPHERD in ranking causal genes stratified by b clinical sites and c primary presenting symptoms. Each boxplot shows the median and interquartile range of the rank of the causal gene. Whiskers extend to ±1.5 × IQR. d Performance of SHEPHERD, six domain-specific approaches, five large language model, traditional machine learning, network science baselines, and a random baseline. The performance metric is average recall at k for k = 1, 5, 10, 25, and 50. Error bars denote standard deviation over models trained with five random seeds. e, f Performance of SHEPHERD against domain-specific algorithms in four extremely hard-to-diagnose scenarios on e EXPERT-CURATED and f VARIANT-FILTERED gene lists. Shown is the win rate, the proportion of patients where SHEPHERD performs the same as or better than the benchmark algorithms.CC BY-NC-ND · PubMed Central
Fig. 4 Causal gene discovery case studies for patients with novel genetic conditions. SHEPHERD identifies the causal gene even in atypical or novel disease presentations. Each patient case study, shown in (a, b), includes the subset of the knowledge graph containing all nodes in the shortest path between the patient’s phenotype terms, causal gene, and disease; a table of the patient’s phenotype terms and attention weights learned by SHEPHERD; and bar plots of scores SHEPHERD assigned to each candidate gene in the EXPERT-CURATED and VARIANT-FILTERED lists. The top and bottom five ranked genes in the VARIANT-FILTERED list are shown. The causal gene is highlighted in orange. The direct phenotypic neighbors of the causal gene are emphasized. In patient UDN-P1's network, the patient’s causal gene is directly connected to the disease in the knowledge graph. In patient UDN-P2's network, there is no disease node because the patient has a novel, uncharacterized syndrome. All panels, except those labeled as a “patient card" (colored box with the information provided by the UDN), depict SHEPHERD's predictions or analyses performed on outputs of SHEPHERD.
Fig. 4 Causal gene discovery case studies for patients with novel genetic conditions. SHEPHERD identifies the causal gene even in atypical or novel disease presentations. Each patient case study, shown in (a, b), includes the subset of the knowledge graph containing all nodes in the shortest path between the patient’s phenotype terms, causal gene, and disease; a table of the patient’s phenotype terms and attention weights learned by SHEPHERD; and bar plots of scores SHEPHERD assigned to each candidate gene in the EXPERT-CURATED and VARIANT-FILTERED lists. The top and bottom five ranked genes in the VARIANT-FILTERED list are shown. The causal gene is highlighted in orange. The direct phenotypic neighbors of the causal gene are emphasized. In patient UDN-P1's network, the patient’s causal gene is directly connected to the disease in the knowledge graph. In patient UDN-P2's network, there is no disease node because the patient has a novel, uncharacterized syndrome. All panels, except those labeled as a “patient card" (colored box with the information provided by the UDN), depict SHEPHERD's predictions or analyses performed on outputs of SHEPHERD.CC BY-NC-ND · PubMed Central
Fig. 5 SHEPHERD identifies patients-like-me from UDN and MyGene2 cohorts. a Performance of SHEPHERD in retrieving MyGene2 patients with the same causal gene as a UDN patient (n = 75 UDN patients with at least one matching patient in the MyGene2 cohort). SHEPHERD is benchmarked against Phrank, a domain-specific algorithm. The performance metric is average recall at k for k = 1, 5, 10, 25, and 50. b Heatmap of the average distance between the phenotype embeddings of pairs of patients across disease categories. Darker colors indicate smaller distances and lighter colors indicate larger distances between patients of each pair of disease categories. c Two-dimensional UMAP plot of SHEPHERD's embedding space of all simulated (circle), UDN (up-facing triangle), and MyGene2 (down-facing triangle) patients colored by their Orphanet disease category. Each of the four case studies consists of a zoomed-in UMAP displaying the query patient (star) and all patients with the same causal gene as the query (colored circles) and a table containing information regarding the top five most similar patients retrieved by SHEPHERD. Patients are bolded in the table if they share the same causal gene. All panels, except those labeled as a “patient card” (colored box with the information provided by the UDN), depict SHEPHERD's predictions or analyses performed on outputs of SHEPHERD.
Fig. 5 SHEPHERD identifies patients-like-me from UDN and MyGene2 cohorts. a Performance of SHEPHERD in retrieving MyGene2 patients with the same causal gene as a UDN patient (n = 75 UDN patients with at least one matching patient in the MyGene2 cohort). SHEPHERD is benchmarked against Phrank, a domain-specific algorithm. The performance metric is average recall at k for k = 1, 5, 10, 25, and 50. b Heatmap of the average distance between the phenotype embeddings of pairs of patients across disease categories. Darker colors indicate smaller distances and lighter colors indicate larger distances between patients of each pair of disease categories. c Two-dimensional UMAP plot of SHEPHERD's embedding space of all simulated (circle), UDN (up-facing triangle), and MyGene2 (down-facing triangle) patients colored by their Orphanet disease category. Each of the four case studies consists of a zoomed-in UMAP displaying the query patient (star) and all patients with the same causal gene as the query (colored circles) and a table containing information regarding the top five most similar patients retrieved by SHEPHERD. Patients are bolded in the table if they share the same causal gene. All panels, except those labeled as a “patient card” (colored box with the information provided by the UDN), depict SHEPHERD's predictions or analyses performed on outputs of SHEPHERD.CC BY-NC-ND · PubMed Central
Fig. 6 SHEPHERD performs novel disease characterization. a Bar plots of the similarity between UDN patients and diseases found in each disease category. We group UDN patients by the disease category of their true disease and show plots for all categories with at least five patients. The bars that do not correspond to the disease category of each patient’s true disease are colored gray. b The column for each of the four case studies contains: the percent similarity distributions of the patient’s phenotype terms to diseases in each disease category based on a phenotype search via the KG (top) or SHEPHERD (bottom), a table of the five most similar diseases according to SHEPHERD, and a table of the patient’s five phenotypic features that are most highly attended by SHEPHERD.
Fig. 6 SHEPHERD performs novel disease characterization. a Bar plots of the similarity between UDN patients and diseases found in each disease category. We group UDN patients by the disease category of their true disease and show plots for all categories with at least five patients. The bars that do not correspond to the disease category of each patient’s true disease are colored gray. b The column for each of the four case studies contains: the percent similarity distributions of the patient’s phenotype terms to diseases in each disease category based on a phenotype search via the KG (top) or SHEPHERD (bottom), a table of the five most similar diseases according to SHEPHERD, and a table of the patient’s five phenotypic features that are most highly attended by SHEPHERD.CC BY-NC-ND · PubMed Central

Verification

Claims the paper makes

20 Valid · 5 Not valid · 6 Not checked · 194 Own results and interpretations

122 more claims · 8 with their cited sources checked

Create a free account to see all 225 claims and to verify the 4 references still open.

See all claims — free account
  • It may state two separate claims as one. 63
  • Probably not a substantial claim. 43
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

3 of 3

Citations

Cites

Show all 127 references

Research works printed but not linked: 6

  1. Brody, S., Alon, U. & Yahav, E. How attentive are graph attention networks? ICLR (2022).
  2. Yang, B., Yih, W.-t., He, X., Gao, J. & Deng, L. Embedding entities and relations for learning and inference in knowledge bases. ICLR (2015).
  3. Goldberger, J., Hinton, G. E., Roweis, S. & Salakhutdinov, R. R. Neighbourhood components analysis. In Proc. 18th International Conference on Neural Information Processing Systems 513–520 (MIT Press, 2004).
  4. Ba, J. L., Kiros, J. R. & Hinton, G. E. Layer normalization. NeurIPS (2016).
  5. Loffe, S. & Szegedy, C. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, 448–456 (2015).
  6. Fey, M. & Lenssen, J. E. Fast graph representation learning with PyTorch Geometric. In ICLR Workshop on Representation Learning on Graphs and Manifolds (2019).

Cited by

See all 54

Metadata

Authors
Alsentzer, E., Li, M. M., Kobren, S. N., Noori, A., Undiagnosed Diseases Network, Kohane, I. S., & Zitnik, M.
Affiliations
  • Harvard University (US)
  • Broad Institute (US)
Venue
npj Digital Medicine
Published
2025-06-20
Volume
8
Issue
1
Pages
380-380
ISSN
2398-6352 (ISSN-L)
OpenAlex
W4411485357
Cited by
54
References
127
Citations by year
2026: 37 · 2025: 12
FWCI
28.59
Type
journal article
Language
en
Open access
Licence
Locations
  • npj Digital Medicine (publishedVersion, cc-by-nc-nd)
  • medRxiv (acceptedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • National Science Foundation (2030459, 2033384)
  • AstraZeneca
  • Microsoft Research
  • Wellcome Trust (WT223718/Z/21/Z)
  • European Bioinformatics Institute
  • National Institutes of Health (U01TR001395, -0001, U01TR002471)
  • U.S. Air Force (FA8702‐15‐D‐0001)
  • Harvard Data Science Initiative, Harvard University
  • National Human Genome Research Institute (U01HG010233, U01HG007709, U01HG007943, U01HG007703, U01HG007530, U01HG010230, U01HG007674, U01HG010217, U01HG007690, U01HG007942, U01HG007708, U01HG010219, U01HG010215, U01HG007672, T32HG002295)
  • National Institute of Neurological Disorders and Stroke (U54NS093793, U54NS108251)
  • National Center for Advancing Translational Sciences
  • Common Fund (U01HG007530, U01TR001395, U01HG007709, U01HG007690, U01HG007672, U01HG007942, U01HG007708, U01HG007943, U54NS093793, U01 HG007703)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Genetics › Genomics and Rare Diseases
  • Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
Keywords
rare genetic disorders, few-shot learning, rare disease diagnosis, knowledge graphs, patient retrieval, Undiagnosed Diseases Network, developmental disorders, phenotypic heterogeneity

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Few shot learning for phenotype-driven diagnosis of patients with rare genetic diseases. https://app.cochat.ai/papers/10.1038/s41746-025-01749-1
The paper
Alsentzer, E., Li, M. M., Kobren, S. N., Noori, A., Undiagnosed Diseases Network, Kohane, I. S., & Zitnik, M. (2025). Few shot learning for phenotype-driven diagnosis of patients with rare genetic diseases. npj Digital Medicine. https://doi.org/10.1038/s41746-025-01749-1

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.