Advancing genetic pathogenicity prediction with three-dimensional proteoform-phenotype analysis
- 1Stanford University
Abstract
Ascertaining the pathogenicity and clinical relevance of genetic mutations is a longstanding challenge across human disease. This is particularly relevant for genes with many variants of unknown significance (VUS), such as PRPH2-retinopathy, where complex variable phenotype patterns and limited genotype correlations hinder application of precision-medicine therapies. Recent AI-based pathogenicity tools, while powerful, may be limited by insufficient mechanistic or phenotypic training data. We established a three-dimensional proteoform-phenotype analysis framework by combining AI-generated protein structural models with clinical genomics data, in which expert consensus adjudicates model outputs through a human-in-the-loop AI approach, to evaluate 46 curated pathogenic missense variants against a background VUS pool. The framework was prospectively validated using an independently evaluated clinical patient cohort. This analysis resolved spatial biophysical signatures corresponding to distinct clinical phenotypes (i.e., retinitis pigmentosa, macular dystrophy, and pattern dystrophy) as well as complex blended presentations. Leveraging these confirmed structural signatures, we prioritized high-risk VUS. This mechanism-resolved framework can be extended to other therapeutically actionable genes to identify novel causal variants and improve diagnosis, prognosis, and better select rare, eligible patients for emerging molecular gene therapy trials.
Plain Language SummaryDistinguishing disease-causing mutations from benign ones remains a major barrier to precision medicine. Here, the authors combine AI-generated protein structures with clinical data and expert human adjudication to show that where a PRPH2 variant sits in three-dimensional space predicts the specific retinal disease it causes, enabling variant prioritization for gene-therapy trials
Related
- AlphaGenome informed machine learning for pathogenicity prediction of germline mutations
- Mechanism-based prediction of insertion-driven high pathogenicity avian influenza virus emergence
- PathoBERT: A Hybrid Attention-Based Genomic Language Model for Read-Level Bacterial Pathogenicity Prediction
- MetaSplice: pathogenicity prediction and ACMG/AMP evidence for intronic variants outside the canonical splice sites
- VitisGP: Multi-Task Benchmark for Genotype to Phenotype Grapevine Prediction
BibTeX
@misc{sun2026advancing,
title = {{Advancing genetic pathogenicity prediction with three-dimensional proteoform-phenotype analysis}},
author = {Y. J. Sun and G. Velez and A. S. Li and A. Kumar and J. T. Vu and C. J. Phillips and D. K. Rasmussen and J. K. Oh and M. G. S. Costa and S. H. Tsang and Y. Sun and A. G. Bassuk and V. B. Mahajan},
howpublished = {medRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.07.26364649},
url = {https://www.medrxiv.org/content/10.64898/2026.10.07.26364649v1},
note = {CC0},
}
RIS
TY - UNPB
TI - Advancing genetic pathogenicity prediction with three-dimensional proteoform-phenotype analysis
AU - Y. J. Sun
AU - G. Velez
AU - A. S. Li
AU - A. Kumar
AU - J. T. Vu
AU - C. J. Phillips
AU - D. K. Rasmussen
AU - J. K. Oh
AU - M. G. S. Costa
AU - S. H. Tsang
AU - Y. Sun
AU - A. G. Bassuk
AU - V. B. Mahajan
PY - 2026
DA - 2026/10/07
PB - medRxiv
DO - 10.64898/2026.10.07.26364649
UR - https://www.medrxiv.org/content/10.64898/2026.10.07.26364649v1
AB - Ascertaining the pathogenicity and clinical relevance of genetic mutations is a longstanding challenge across human disease. This is particularly relevant for genes with many variants of unknown significance (VUS), such as PRPH2-retinopathy, where complex variable phenotype patterns and limited genotype correlations hinder application of precision-medicine therapies. Recent AI-based pathogenicity tools, while powerful, may be limited by insufficient mechanistic or phenotypic training data. We established a three-dimensional proteoform-phenotype analysis framework by combining AI-generated protein structural models with clinical genomics data, in which expert consensus adjudicates model outputs through a human-in-the-loop AI approach, to evaluate 46 curated pathogenic missense variants against a background VUS pool. The framework was prospectively validated using an independently evaluated clinical patient cohort. This analysis resolved spatial biophysical signatures corresponding to distinct clinical phenotypes (i.e., retinitis pigmentosa, macular dystrophy, and pattern dystrophy) as well as complex blended presentations. Leveraging these confirmed structural signatures, we prioritized high-risk VUS. This mechanism-resolved framework can be extended to other therapeutically actionable genes to identify novel causal variants and improve diagnosis, prognosis, and better select rare, eligible patients for emerging molecular gene therapy trials. Plain Language SummaryDistinguishing disease-causing mutations from benign ones remains a major barrier to precision medicine. Here, the authors combine AI-generated protein structures with clinical data and expert human adjudication to show that where a PRPH2 variant sits in three-dimensional space predicts the specific retinal disease it causes, enabling variant prioritization for gene-therapy trials
ER -