Symptom co-occurrence modeling for phenotype-driven diagnosis in genetic diseases

Phenotype-driven gene prioritisation relies on matching the Human Phenotype Ontology (HPO) terms recorded by clinicians to phenotype-gene associations held in knowledge resources. How physicians phenotype patients varies widely, and this heterogeneity is poorly captured by existing resources, limiting diagnostic performance. We characterised phenotyping practice in 1,686 patients with a confirmed genetic diagnosis from four international cohorts, and in a prospective experiment in which 12 clinical geneticists encoded the same three clinical reports in HPO terms. We built PhenoGenius, a symptom co-occurrence model derived from phenotype-gene associations collected from HPO-structured databases and from text mining of OMIM, MedGen and PubMed, combining a node-similarity graph and non-negative matrix factorisation. We benchmarked PhenoGenius against seven phenotype-driven gene prioritisation tools and two reasoning language models (OpenAI o3 and Gemini 2.5 Pro), and evaluated it on 4,716 curated case reports from the GA4GH Phenopacket Store. Only 43% of the 11,526 gene-phenotype associations recorded in the cohort were present in HPO-structured databases. Among patients sharing the same causal gene, only 22.7% of annotations were reused even when any ancestor or descendant term was allowed; when different physicians encoded the same report, 47.3% of annotations had no ontological counterpart in another physician’s rendering. The model summarised 16,600 HPO terms into 390 groups of co-occurring symptoms and covered 11,340 (98.4%) of the cohort associations. Of 28,862 associations added to the HPO between 2022 and 2025, 34.8% had been predicted by the factorisation in 2022. PhenoGenius reduced the median rank of the causal gene from 65 with the best comparator (CADA) to 35, an advantage reproduced in each of the four cohorts, and recovered the causal gene within the top 500 in 87.0% of patients, against 81.6%. On curated case reports, the causal gene ranked within the top 10 in 48.6% of cases. Symptom co-occurrence modelling captures the variability of real-world phenotyping and extends the coverage of phenotype-driven gene prioritisation, offering a scalable approach to support genomic diagnosis and precision medicine.

Authors

Institutions

Publication Details

Journal
Genome Medicine
Published
2026-11-14
DOI
https://doi.org/10.1186/s13073-026-01828-2
Primary Topic
Biomedical Text Mining and Ontologies
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Symptom co-occurrence modeling for phenotype-driven diagnosis in genetic diseases

Radu Harbuz, Sacha Beaumeunier, Quentin Testard, Jérôme Audoux et al.
Genome Medicine
Biomedical Text Mining and Ontologies
article

Symptom co-occurrence modeling for phenotype-driven diagnosis in genetic diseases

Radu Harbuz, Sacha Beaumeunier, Quentin Testard, Jérôme Audoux, Florence Amblard, Jean‐Marie Ravel, DImitri Larue, Michael S. Blum, Mouna Barat‐Houari, Nicolas Duforet-Frebourg, Véronique Satre, Wilfrid Carré, Charles Coutton, Virginie Bernard, Pauline Le Tanno, Julien Thévenon, Marjolaine Willems, Laure Raymond, Françoise Devillard, Cédric Le Maréchal, Damien Sanlaville, David Geneviève, Kévin Yauy, Nicolas Chatron, Laurent Mesnard, Nicolas Philippe, Yannis Duffourd, Sylvie Odent, Denis Bertrand, Isabelle Marey, Marc Planes, Claire Bardel, Laurence Faivre, Klaus Dietrich, Alexandre Boulat, Benoit Simard, Jiri Ruzicka, Pierre Ray, Marie-Pierre Audrezet, Marie de Tayrac, Thomas Guignard, Jean-François Taly, Frederic Tran Mau-Them, Clément Pavie Routaboul
article en

Abstract

Phenotype-driven gene prioritisation relies on matching the Human Phenotype Ontology (HPO) terms recorded by clinicians to phenotype-gene associations held in knowledge resources. How physicians phenotype patients varies widely, and this heterogeneity is poorly captured by existing resources, limiting diagnostic performance. We characterised phenotyping practice in 1,686 patients with a confirmed genetic diagnosis from four international cohorts, and in a prospective experiment in which 12 clinical geneticists encoded the same three clinical reports in HPO terms. We built PhenoGenius, a symptom co-occurrence model derived from phenotype-gene associations collected from HPO-structured databases and from text mining of OMIM, MedGen and PubMed, combining a node-similarity graph and non-negative matrix factorisation. We benchmarked PhenoGenius against seven phenotype-driven gene prioritisation tools and two reasoning language models (OpenAI o3 and Gemini 2.5 Pro), and evaluated it on 4,716 curated case reports from the GA4GH Phenopacket Store. Only 43% of the 11,526 gene-phenotype associations recorded in the cohort were present in HPO-structured databases. Among patients sharing the same causal gene, only 22.7% of annotations were reused even when any ancestor or descendant term was allowed; when different physicians encoded the same report, 47.3% of annotations had no ontological counterpart in another physician’s rendering. The model summarised 16,600 HPO terms into 390 groups of co-occurring symptoms and covered 11,340 (98.4%) of the cohort associations. Of 28,862 associations added to the HPO between 2022 and 2025, 34.8% had been predicted by the factorisation in 2022. PhenoGenius reduced the median rank of the causal gene from 65 with the best comparator (CADA) to 35, an advantage reproduced in each of the four cohorts, and recovered the causal gene within the top 500 in 87.0% of patients, against 81.6%. On curated case reports, the causal gene ranked within the top 10 in 48.6% of cases. Symptom co-occurrence modelling captures the variability of real-world phenotyping and extends the coverage of phenotype-driven gene prioritisation, offering a scalable approach to support genomic diagnosis and precision medicine.

Genome Medicine
Centre National de la Recherche Scientifique (FR), Inserm (FR), Université de Montpellier (FR), Centre Hospitalier Universitaire de Grenoble (FR), Centre Hospitalier Universitaire de Montpellier (FR), Institute for Regenerative Medicine & Biotherapy (FR), Laboratoire d'Informatique, de Robotique et de Microélectronique de Montpellier (FR), Institut pour l'avancée des biosciences (FR), Nutrition-Génétique et Exposition aux Risques Environnementaux (FR), Université Grenoble Alpes (FR), Université de Lorraine (FR)
Good health and well-being
Biomedical Text Mining and Ontologies
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon