Revolutionizing rare disease diagnosis: Houston's AI breakthrough

Genetic diseases arising from a mutation in a single gene are often called Mendelian or monogenically inherited diseases. The term "Mendelian" is derived from Gregor Mendel, who studied inheritance patterns and the transmission of traits using green peas.

Diagnosing rare Mendelian disorders is challenging, even for skilled geneticists. Thus, scientists at Baylor College of Medicine aim to simplify this process through the use of AI. The AI-MARRVEL (AIM) machine learning system helps identify the most likely disorder variants.

Dr. Pengfei Liu, the Clinical Deputy Director at Baylor Genetics, mentioned that "the recognition rate of rare genetic diseases is only about 30 percent, and it takes an average of six years from the first symptoms to get a diagnosis. There's a pressing need for new methods to improve the speed and accuracy of diagnosis."

AIM is trained on a public database of known variants and genetic analysis called Model organism Aggregated Resources for Rare Variant Exploration (MARRVEL), developed earlier by the Baylor team. The MARRVEL database contains over 3.5 million variants from thousands of diagnosed cases.

When scientists input data about gene sequences and patient symptoms into AIM, it ranks the most likely genes causing a rare disease. Scientists have compared AIM's performance with other algorithms on three data sets with confirmed diagnoses from Baylor Genetics, the Undiagnosed Diseases Network funded by the National Institutes of Health (UDN), and the Deciphering Developmental Disorders (DDD) project. AIM consistently placed diagnosed genes as the top candidate twice as often as all other methods tested against these real-world data sets.

Dr. Zhandong Liu from Baylor, a study co-author, stated: "We trained AIM to mimic human decision-making processes, and the machine can outperform humans in terms of speed, efficiency, and cost. This approach effectively doubled the diagnostic accuracy rate."

In rare diseases that have remained unsolved for years, AIM offers new hope, enabling reanalysis with the latest knowledge. "By using AIM to identify a set of high-confidence cases that can be resolved and then manually reviewed, we can significantly streamline the reanalysis process," Zhandong Liu explained. "We expect this tool could uncover an unprecedented number of cases previously considered undiagnosable."

AIM's capability to discover new gene candidates not previously linked to any disease was also tested. It accurately predicted two newly reported disease genes as the main "suspects" in two instances.

The team at Baylor plans to develop the next generation of diagnostic intelligence and implement it in clinical settings.