Backlog of unresolved medical diagnoses leads to new method for clinical exome reanalysis

Child Hospital Social

In a step toward efficient and cost-effective novel disease gene discovery, experts at Texas Children's Hospital and Baylor College of Medicine (BCM) have developed an automated way to reevaluate unsolved medical cases and potentially pinpoint genetic diseases that have recently become diagnosable.

Drawing from expertise in molecular and human genetics, quantitative and computational biomedical sciences, data science and software engineering in pediatrics and neurological research, AI-MARRVEL (AIM) emerged this year as a new artificial intelligence (AI) system of analysis that prioritizes causative genes and variants for Mendelian disorders based on patients’ clinical features and exome sequencing profiles. A study of AIM was published April 25 in New England Journal of Medicine AI (NEJM AI).

Updates to curated variants in disease databases afforded the opportunity to reevaluate a set of previously unsolved cases, according to first author Dr. Dongxue Mao and colleagues from Baylor Genetics, Texas Children's, and the Jan and Dan Duncan Neurological Research Institute in Houston. While the work is far from complete, researchers from the Baylor Genetics clinical diagnostic laboratory noted that AIM can contribute to predictions independent of clinical knowledge of the gene of interest, helping to advance the discovery of novel disease mechanisms.

Using unsolved pools that have accumulated over time, the team designed a confidence metric on which their software achieved a precision rate of 98% and identified more than half of diagnosable cases out of a collection of 871.

Diagnosing genetic disorders

Mendelian genetic disorders are in the rare disease category and can be severe. Those affected may undergo numerous medical procedures over the course of their lives but never arrive at a single diagnosis. Mendelian disorders affect more than 400 million people worldwide, according to Baylor, a clinical site designated among several Centers for Mendelian Genomics by the National Institutes of Health (NIH).

Numbering in the thousands, Mendelian disorders are generally thought to be caused by mutations in a single gene. These conditions create significant hardships in life and along the healthcare continuum. Commonly known examples of Mendelian disorders include sickle cell anemia, muscular dystrophy, cystic fibrosis, thalassemia, phenylketonuria (aka PKU), color blindness, skeletal dysplasia, and hemophilia for example, but many more are less known and understood.

“The diagnostic rate for rare genetic disorders is only about 30%, and on average, it is six years from the time of symptom onset to diagnosis. There is an urgent need for new approaches to enhance the speed and accuracy of diagnosis,” said co-corresponding author Dr. Pengfei Liu, associate professor of molecular and human genetics and associate clinical director at Baylor Genetics, for a BCM blog post.


The foundation of the AI is MARRVEL, which stands for Model organism Aggregated Resources for Rare Variant ExpLoration. Developed by the Baylor team, MARRVEL is a public database of more than 3.5 million known variants and genetic analyses from thousands of diagnosed cases.

How it works: researchers provide the AIM algorithm with patients' exome sequence data and symptoms, and AIM ranks diagnostic variants into the most likely gene candidates causing the rare disease.

The algorithm used an initial random-forest machine-learning (ML) classifier provided by scikit-learn and trained on MARRVEL's collection. In addition, the training samples relied on exome sequencing, rather than whole-genome sequencing, because exome sequencing can detect a wider array of variants that include deep intronic, copy-number variations, and structural variations, according to the researchers.

However, AIM currently is not equipped to analyze structural variations or copy-number variations. Furthermore, the authors wrote, "AIM has been predominantly trained on cases with coding variants, which constrains its capacity to effectively prioritize noncoding variants." AIM can can process single-nucleotide variants and small insertions or deletions.

Exome sequencing data and human phenotype ontology (HPO) terms were compiled from three distinct real-world patient groups: the Clinical Diagnostic Lab, Undiagnosed Disease Network (UDN), and the Deciphering Developmental Disorders (DDD) project.

In an announcement celebrating progress, Baylor researchers said the AI has diagnosed patients from the three independent cohorts and been compared to other algorithms designed and used for similar purposes.

Backlog of cold cases

"Hundreds of novel disease-causing variants that may be key to solving these cold cases are reported every year; however, determining which cases warrant reanalysis is challenging because of the high volume of cases," stated Baylor's Dr. Zhandong Liu, chief computational scientist at Texas Children's Hospital. "The researchers tested AIM’s clinical exome reanalysis on a dataset of UDN and DDD cases and found that it was able to correctly identify 57% of diagnosable cases," Liu explained.

Liu continued, “We can make the reanalysis process much more efficient by using AIM to identify a high confidence set of potentially solvable cases and pushing those cases for manual review. We anticipate that this tool can recover an unprecedented number of cases that were not previously thought to be diagnosable.”

AI uses feature climbing method

To better understand how AIM arrives at its predictions, researchers developed what they call a "feature climbing" method to evaluate the contribution of each feature by perturbing the feature value and re-running the predictions, Mao and colleagues stated in the NEJM AI article.

As part of AIM's knowledge-based feature engineering, 56 raw features were selected, encompassing disease database, minor allele frequency, variant impact, evolutionary conservation, inheritance pattern, phenotype matching, gene constraint, variant pathogenicity prediction scores, splicing prediction, and sequencing quality. The team then generated 47 additional features.

AIM performs using six different modules that incorporate prior learning. Module 1, for example, evaluates whether the candidate variant or corresponding gene is curated in disease databases such as OMIM, ClinVar, or others. Each module offers an independent analysis of the pathogenicity of a variant or gene, producing new features that emulate the decision-making of human experts, according to the article in NEJM AI.

"We quantify the effect size of each feature as the maximum difference between the prediction score and the minimum value achievable through perturbation," the authors wrote. "All features are grouped into different classes based on their biological meaning. The biggest effect size was seen when perturbing the minor allele frequency, followed by the variant curation status in disease databases and phenotypic matching."

Automated analysis

For the BCM blog, Dr. Fan Xia commented that combined with the expertise of certified clinical laboratory directors, highly curated datasets and automated technology such as AIM establishes a form of augmented intelligence to provide "comprehensive genetic insights at scale, even for the most vulnerable patient populations and complex conditions." Xia is associate professor of molecular and human genetics at Baylor and vice president of clinical genomics at Baylor Genetics.

AIM was trained using samples that were clinically diagnosed and curated by American Board of Medical Genetics and Genomics-certified experts, along with additional engineered features that encode prior knowledge, such as genetic principles and the knowledge of clinical genetics experts. At present, the AIM interface employs ClinPhen for mapping clinical notes into phenotype terms. Yet, large language models (LLMs) such as PhenoBCBERT49 and PhenoGPT49 have shown "superior" performance, the researchers noted, adding that the LLMs may be considered for future integration into the AIM platform.

For other researchers, AIM has a Web interface (available at that enables users to submit cases and interactively review the results. The interface provides automatic extraction of HPO terms from clinical notes using ClinPhen, and users have the flexibility to refine the extracted HPO terms through ontology trees, according to the AIM team.

To see how the developers used inheritance information, molecular evidence, and phenotypic data in diagnostic performance, and more, read the full article here.

This work was supported by the Chan Zuckerberg Initiative, the National Institute of Neurological Disorders and Stroke (3U2CNS132415), the Chao Endowment, and the Huffington Foundation.

Page 1 of 17
Next Page