Artificial Intelligence Aiding Human Interpretation Of Sequenced Genomes
By Deborah Borfitz
October 28, 2021 | A “quiet revolution” has been underway over the past year, exponentially growing the number of genetic variants discoverable by genome sequencing, according to Stephen Kingsmore, M.D., president and CEO of Rady Children's Institute for Genomic Medicine in San Diego. Notably, it is now possible to identify an enormous number of structural variants—large genomic alterations that include insertions, deletions, and copy number variations—which account for about 20% of diagnoses in the neonatal intensive care unit (NICU).
When whole genome sequencing (WGS) was first introduced about a decade ago, the technology was limited to single nucleotide variants and short indel (insertions/deletion) variants, says Kingsmore. “We’re now recognizing that there are literally tens of thousands of events that have never before been seen in a genome but offer the potential for us to significantly improve our diagnostic rate… [and] we are just scratching the surface.”
Larger structural variants eclipse the short variants scientists were focused on for years, in terms of the number of affected bases in a DNA sequence. Chromosomal microarray analysis was traditionally used, which captured only “really big events” while completely missing smaller changes.
Physicians have been accepting the idea that only about one-third of their pediatric patients with a suspected genetic disease would be diagnosable, he continues. Whether the deluge of genetic information increases the diagnostic yield by 5% or 15% has yet to be determined in clinical studies.
Ensuring no child or family waits for a diagnosis will require a heavy investment in automation and artificial intelligence (AI) to make sense of all the newly available data, Kingsmore says. Geneticists can’t decipher the meaning of identified variants fast enough to cope with the ramp-up in children who need testing. “The interpretation cost has now become the largest single cost in terms of genomic reports.”
WGS has predominantly been a tool of research rather than routine clinical care, says Kingsmore. Costing thousands of dollars per test, its use until recently has been highly dependent on grants and philanthropic funding.
Over the past five years, however, the number of hospitals across North America regularly sending blood samples to Rady Children’s nonprofit Institute for rapid WGS has grown from zero to 72, he reports. “It is starting to become recognized as a first-tier test in diagnostic evaluation.”
Demand for WGS continues to rise each month and much of it of late is originating in Michigan, says Kingsmore. As of September 1, Michigan became the first state to make WGS a covered benefit for eligible infants enrolled in Medicaid and reimburses hospitals with a “carve out” payment rather than as part of an existing diagnosis-related group (DRG).
Previously, Spectrum Health was the only payer in the state covering the test, he notes. The addition of Michigan Medicaid is big news, given that the joint federal-state program pays the bills of 50% to 60% of babies admitted to the NICU.
In California, meanwhile, a legislative bill was signed by the governor in July making WGS a covered benefit for Medi-Cal babies. But the law has not yet translated into a policy, says Kingsmore. “If California considers this to be part of a DRG, it will not move the dial one iota because there is no way hospitals can cover the cost of this test out of pocket because it is still expensive.”
Decision Support Systems
The latest advance, as recently reported in Genome Medicine (DOI: 10.1186/s13073-021-00965-0), involves using AI to expedite genome interpretation and nominate candidate diagnoses for rare genetic diseases. Researchers (including Kingsmore) were able to detect more than 90% of disease-causing variants in infants with rare diseases across six genomic centers and hospitals using Fabric GEM—a newly released, AI-based clinical decision support tool—and whole-genome and whole-exome data from previously diagnosed newborns and rare disease patients.
To benchmark GEM’s performance, the study collected 119 clinical cases from Rady Children’s Hospital and another 60 from five other clinical sites: University of Utah Health, Boston Children’s Hospital, Christian-Albrechts University of Kiel & University Hospital Schleswig-Holstein (Germany), Tartu University Hospital (Estonia), and the Translational Genomics Research Institute.
Fabric Genomics worked with Mark Yandell, Ph.D., a founding scientific advisor to the company as well as professor of human genetics and Edna Benning Presidential Endowed Chair at U of U Health, to arrive at the “winning formula,” says Kingsmore. Cases for training the algorithm were submitted in batches by Rady Children’s before being validated across institutions.
The decision support tool is agnostic to the way cases get collected and the sequencing methods used because diagnostic laboratories around the world tend to follow the same protocols, he says. The genome data is therefore easy to harmonize.
GEM is the latest major addition to the analysis platform of Fabric Genomics, and the first of about 10 commercial software solutions that have emerged over the last decade and begun offering AI-assisted interpretation of structural variants, says Kingsmore. Rady Children's Institute for Genomic Medicine adopted the platform at its establishment six years ago.
The available software systems are all somewhat similar and, much like an electronic health record (EHR), the one adopted tends to become a long-term commitment, Kingsmore notes. The alternatives include Illumina, which has built software the Institute uses to identify variants in genomes.
In the U.K., the players include Genuity Science (originally Amgen-owned deCODE Genetics) that recently cut off its Chinese operations. The British government also owns its own company, Genomics England, and is using it to introduce sequencing technology to mainstream healthcare.
All clinical groups initially built their own in-house system and some, including Baylor Genetics and GeneDx, continue to support them, he says. Stanford helped develop an open-source system for processing WGS data a few years ago and more recently launched a comprehensive an-house service for WGS.
GEM Method
The idea behind GEM software is to help clinical geneticists arrive at a diagnosis by having the AI tool analyze data from clinical notes, medical databases, and genome sequences specific to one patient and, based on its training on the same sort of data, find associations with the more than 6,700 identified disease-associated genes.
GEM comprises two main AI algorithms already being used by human interpreters—Phevor, which recognizes the clinical features of genetic diseases and rank-orders them based on the phenotypic characteristics observed in the child, and VAAST, which does the same with the variants, Kingsmore explains. The variants get ranked in terms of how likely they are to cause disease, from “benign… to definitely disease-causing.”
As demonstrated in the recent study, natural language processing (NLP), can also be used as an add-on to GEM for the extraction of key clinical features from the EHR, eliminating another manual step. For this, the researchers used the CLiX ENRICH software of Clinithink.
The chain of events is “pretty straightforward,” says Kingsmore. After a physician orders genomic testing, the patient’s blood sample gets genome-decoded, likely uncovering a few million variants. Meanwhile, NLP goes into the EHR and grabs the documented symptoms. “Those are the two data streams that then converge in GEM to be analyzed and give a diagnosis.”
The diagnostic sensitivity of GEM is similar to that of AI-based software called Moon developed by Diploid (Belgium), he says. The privately held company was acquired last year by the medical genetics company Invitae.
“Moon works very, very well, but it is a black box and that was something that lab directors were quite reluctant to embrace,” notes Kingsmore.
Machine Limitations
Although GEM is completely automated, producing results with the literal click of a button, it is currently used primarily as an adjunct to manual interpretation, Kingsmore says. GEM speeds up the process because it considers variant quality, deleteriousness, prior clinical annotations, and mode of inheritance. GEM gene scores also summarize the relative strength of evidence suggesting a gene is damaged and can explain the observed phenotype.
One of the attractions of the AI software is that it is “not a pureplay black box,” says Kingsmore, since GEM combines the scores from Phevor and VAAST into the final rankings. Users are accustomed to seeing scores from those algorithms and like having the ability to “dial [them] up or dial down to look at a broader or narrower set of possibilities.”
They might opt to look at the state and frequency of variants or certain genes of interest. Aided by GEM, genome interpretation can be performed in as little as an hour. That compares to an average of a day in the absence of the tool.
Manual interpretation is unlikely to be abandoned any time soon, says Kingsmore. “[GEM] is brand new, so people will need to gain confidence in its ability. In vitro diagnostics is also heavily regulated, so people want to be sure it conforms to their compliance and quality assurance protocols and various licenses.” That process will probably take about 18 months.
Furthermore, “AI can’t solve it all,” he adds. “There will always be a subset of cases where either the software gets it wrong or misses something.” Until the decision support tool is “fairly perfect,” some labs will be uncomfortable fully relying on its output.
“We are still in the infancy of the genomic era… [and] our knowledge of these genetic diseases is still very new,” Kingsmore points out. From that perspective, GEM’s 90% accuracy rate might be viewed as pleasantly surprising—and certain only to get better with time.
“This is all moving very fast,” he adds. “We have discovered 600 new genetic diseases in the last four to five years.”
Vast Potential
So far this year, Gem has been used on about 1,000 cases at Rady Children’s Institute for Genomic Medicine and lab directors there have found it to be useful, says Kingsmore. Knowing the symptoms is vital to its use in guiding diagnosis.
As was recently shared in the New England Journal of Medicine (DOI: 10.1056/NEJMc2100365), a 13 and a half hour genome study at Rady Children’s (aided by Illumina), saved the life of a 42-day-old infant whose older sister had died a decade earlier from the same rare, but treatable, genetic condition—called autosomal recessive thiamine metabolism dysfunction, syndrome 2. Within six hours of administering high doses of thiamine and biotin, the child’s condition stabilized, and he was discharged a few days later. (Six years ago, Kingsmore made news with 26-hour whole genome diagnostic sequencing.)
“It is a beautiful example of what we want to do all over the country for every child who might have a treatable condition,” says Kingsmore.
The utility of WGS and associated clinical decision support tools to analyze the genome of healthy individuals and detect problems is a “much tougher proposition,” he adds. “Eventually we will know the bulk of the type of mutations that cause disease and [whole genome sequencing] will become more of a database lookup [exercise] we will be able to do at birth.”
Currently, the focus of Rady Children's Institute for Genomic Medicine is better diagnosing hospitalized children with diseases of unknown etiology, most of whom are in either the neonatal, pediatric, or cardiac ICU, says Kingsmore. Next up will be outpatient genomic testing, which doesn’t have the same urgency for rapid results and would therefore be less expensive.
Other non-pediatric applications of the technology are under investigation by researchers worldwide. In a small study on 50 adult patients with acute cardiovascular events (mostly heart attacks) published last year in Circulation: Genomic and Precision Medicine (DOI: 10.1161/CIRCGEN.120.002961), researchers from the Institute and Brigham and Women’s Hospital found that rapid WGS could identify pathogenic variants in one out of five patients.
Much work remains to be done on that front, Kingsmore says, but rapid WGS is going to “gradually percolate throughout medicine.” Newborns are only the starting point.
In a forthcoming paper, researchers at the Institute will address how to optimize an NLP tool for the EHR a health system is using, he says. Rady Children’s, which uses Epic, has done this for two NLP methods it favors.
Work on how to efficiently extract data out of the EHR is ongoing, he adds. The systems were originally built for billing and are consequently imperfect for informing diagnosis, but they’re slowly getting better.