Artificial intelligence (AI) models trained on genomic, pathology, and other biological data could help improve the diagnosis of rare diseases by supporting variant interpretation, identifying disease-causing mutations, and assisting with treatment selection, according to an article published in the Journal of Medical Internet Research.
An estimated 446 million people worldwide are living with a rare disease, yet fewer than five percent of these conditions have an approved treatment. Many patients also experience years of diagnostic uncertainty, even after genomic sequencing. The article reviews several emerging AI tools designed to help laboratories and clinicians interpret increasingly complex genomic and pathology data, while noting that these technologies require further clinical validation before routine use.
One of the biggest diagnostic challenges is interpreting genomic sequencing results. A patient's genome may contain more than 10,000 protein-coding variants, but only a small number are likely to contribute to disease. Distinguishing pathogenic variants from benign findings remains a major obstacle in rare disease diagnosis.
The article highlights popEVE, an AI model developed to predict which genetic variants are most likely to be disease-causing. By combining protein language modeling with population genetic data, the model identified 123 previously unrecognized variants associated with severe developmental disorders. More than 25 of the implicated genes have since been independently confirmed and added to disease-gene databases. Researchers are now working to expand the approach to identify combinations of variants that affect the same biological pathway.
The review also describes AI systems designed to assist, rather than replace, clinical decision-making. DeepRare, for example, was evaluated using nine rare disease datasets and correctly identified the most likely diagnosis in 64 percent of cases. In a separate study, researchers used OpenAI's o3 Deep Research model to analyze genomic data from 376 previously unsolved rare disease cases. Following expert review and additional laboratory testing, 18 patients received a confirmed molecular diagnosis.
Foundation AI models trained on large biological datasets may also have applications in pathology. The article highlights TITAN, a model trained on more than 330,000 digitized biopsy images that generated reports to support the diagnosis of rare cancers and performed better than existing diagnostic aids. Other foundation models are being developed to predict tumor behavior, treatment response, and potential drug-repurposing opportunities for rare diseases.
Despite encouraging early results, the authors note that AI performance depends on high-quality, representative datasets. Current models may be affected by incomplete or biased training data, and their ability to support diagnosis, biomarker discovery, and treatment selection will require prospective clinical validation before widespread adoption.
These technologies may ultimately provide additional decision-support tools for prioritizing genetic variants, interpreting molecular and histopathology findings, and accelerating rare disease diagnosis. The authors emphasize, however, that AI is intended to complement – not replace – clinical and laboratory expertise.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
