ZebraSeek integrates complementary facial and clinical evidence for rare disease prioritization
Abstract
Rare diseases can remain undiagnosed for years because individual disorders are uncom- mon and clinically heterogeneous, while relevant clues are distributed across structured phenotypes, facial appearance, prior cases and disease knowledge. Computational tools can use Human Phenotype Ontology (HPO)-encoded findings, genomic data and increas- ingly diverse medical images, but these modalities are processed by models with different outputs and score scales. ZebraSeek converts clinical and facial inputs into ranked disease candidates with specialist tools, then uses a large language model (LLM) to integrate and externally check those candidates while retaining their sources. Facial-image analysis provides a complementary machine-readable phenotype that can capture patterns not fully represented by manually encoded HPO terms. In 74 literature-derived cases spanning 19 diseases, ZebraSeek placed the recorded diagnosis first in 50 (67.6%) and within five in 59 (79.7%); it recovered 11 diagnoses outside PubCaseFinder’s top five, while five diagnoses present in PubCaseFinder’s top five were absent from ZebraSeek’s final top five. During DBCLS BioHackathon 2026, we expanded analysis to 368 patients, 462 images and 54 disorders to examine how candidate depth affects coverage and downstream candidate count, and to assess the feasibility of case-specific candidate selection.
Related
- miRAssist: a context-aware, evidence integration framework for interpretable miRNA-target prioritization
- Dated public evidence and known genes for retrospective evaluation of gene prioritization
- Complementary axes of functional system integrity and integration with divergent clinical correlates characterize focal epilepsy
- Evidence for decoder-label-regime-dependent effects of complementary leaky integrate-and-fire temporal-state decoding in motor-imagery EEG
- Global use of traditional, complementary and integrative medicine among adults with cancer in hospital and clinical settings: A systematic review and meta-analysis
BibTeX
@misc{yoshikuwa2026zebraseek,
title = {{ZebraSeek integrates complementary facial and clinical evidence for rare disease prioritization}},
author = {Naoya Yoshikuwa and Hirokazu Chiba and Teppei Okazaki and Jaemoon Shin and Eisuke Dohi and Hiroyuki Mishima and Atsuko Yamaguchi and Tzung-Chien Hsieh and Orion Buske and Susumu Goto},
howpublished = {BioHackrXiv},
year = {2026},
month = oct,
doi = {10.37044/osf.io/kxqb7_v1},
url = {https://osf.io/preprints/biohackrxiv/kxqb7_v1/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - ZebraSeek integrates complementary facial and clinical evidence for rare disease prioritization
AU - Naoya Yoshikuwa
AU - Hirokazu Chiba
AU - Teppei Okazaki
AU - Jaemoon Shin
AU - Eisuke Dohi
AU - Hiroyuki Mishima
AU - Atsuko Yamaguchi
AU - Tzung-Chien Hsieh
AU - Orion Buske
AU - Susumu Goto
PY - 2026
DA - 2026/10/16
PB - BioHackrXiv
DO - 10.37044/osf.io/kxqb7_v1
UR - https://osf.io/preprints/biohackrxiv/kxqb7_v1/
AB - Rare diseases can remain undiagnosed for years because individual disorders are uncom- mon and clinically heterogeneous, while relevant clues are distributed across structured phenotypes, facial appearance, prior cases and disease knowledge. Computational tools can use Human Phenotype Ontology (HPO)-encoded findings, genomic data and increas- ingly diverse medical images, but these modalities are processed by models with different outputs and score scales. ZebraSeek converts clinical and facial inputs into ranked disease candidates with specialist tools, then uses a large language model (LLM) to integrate and externally check those candidates while retaining their sources. Facial-image analysis provides a complementary machine-readable phenotype that can capture patterns not fully represented by manually encoded HPO terms. In 74 literature-derived cases spanning 19 diseases, ZebraSeek placed the recorded diagnosis first in 50 (67.6%) and within five in 59 (79.7%); it recovered 11 diagnoses outside PubCaseFinder’s top five, while five diagnoses present in PubCaseFinder’s top five were absent from ZebraSeek’s final top five. During DBCLS BioHackathon 2026, we expanded analysis to 368 patients, 462 images and 54 disorders to examine how candidate depth affects coverage and downstream candidate count, and to assess the feasibility of case-specific candidate selection.
ER -