Over the past two decades, researchers have made significant advances in genomics, allowing them to identify more than 7,000 rare genetic diseases. Despite these impressive advances in genomic technology and variant interpretation pipelines, more than half of rare disease cases remain unsolved. The main reason for this is that many genetic disease associations are still unknown or extremely rare. As a result, many patients wait years for a diagnosis. During this time, symptoms may progress and hope may begin to fade. For people living with rare diseases, timely and accurate diagnosis can have a transformative effect on guiding treatment, reducing uncertainty, and providing peace of mind.
To address this challenge, Genome-Wide Sequencing Ontario (GSO), a collaboration between the Hospital for Sick Children (SickKids) and the Children’s Hospital of Eastern Ontario (CHEO), is leveraging Illumina technology and expertise. GSO is funded by the Ontario Ministry of Health to provide whole-genome clinical sequencing services to Ontarians with suspected rare diseases. Since the start of 2021, the target audience has expanded and testing demand has increased significantly, necessitating more efficient analysis while maintaining quality. “For patients with rare diseases, it can take years to receive a diagnosis,” says Christian Marshall, director of the Clinical Molecular Laboratory in SickKids’ genomic diagnostics division. “This is partially due to the inability to interpret many of the genome’s rare variants, which is further exacerbated by manual analysis and reporting bottlenecks. Improving variant annotation pipelines and using new tools are essential to delivering accurate and timely results to patients.”
Evaluate AI to provide faster and more accurate diagnosis
However, improving these pipelines can be a time-consuming and labor-intensive process that is not sustainable at scale using human analysts. The analysis workflow has multiple steps, from variant calling, which takes a patient’s genome and compares it to a reference genome to find variants, to variant annotation, which uses known databases to add context to variants. This information is used to interpret the potential pathogenicity of the variant. This process has traditionally been done manually.
In hopes of having a real impact on diagnostic yield and turnaround time, the team decided to evaluate the application of artificial intelligence (AI) to variant analysis workflows. “Even with annotation software, it can be very difficult to understand whether a variant is pathogenic or not,” says Livia Loureiro, senior staff medical scientist in Illumina’s medical affairs division. “A single genome contains approximately 4 to 5 million mutations. Genome sequencing captures a wide range of variations, including single nucleotide variations, structural variations, and short tandem repeats, making the data richer but much more complex to interpret. Analyzing data quickly and accurately at that scale is not possible without adding AI to the workflow.”
The team conducted a retrospective study using 852 pediatric cases with known causative variants using explainable AI variant interpretation software for secondary and tertiary analyses. They evaluated the new AI model and compared it to previous versions. They found that both AI versions were successful in prioritizing causative mutations in patient cases, with version 1 ranking 98.3% and version 2 ranking 98.8% at the top (‘most likely’). Considering all prioritized predictions, both models achieved 99% capture of causal variants. We also found that the new AI version had improved specificity, detecting 100% of the top 10 reported variants, compared to the older version, which detected 94% of variants.
“This is an important milestone,” Loureiro said. “Evidence supports the use of AI-powered variant analysis to prioritize variants quickly and with unprecedented accuracy in a large number of cases.”
Marshall hopes that AI can go a step further than variant interpretation and support the collection of structured phenotypes from clinical records to further improve the accuracy of variant interpretation. Of course, some degree of human oversight is still required, especially for variants with low penetrance or with questionable sequence or mapping quality. “AI is poised to have a major impact on rare disease diagnosis,” Marshall says. “However, we also found that detailed human evaluation is required before full integration. The key, and the main challenge, is to design robust validation and use it responsibly.”
More answers are possible when data is shared
To expand the potential impact of AI variant interpretation, Loureiro and Marshall are also developing a shared data network among Canadian laboratories. Because rare diseases affect so few people, clinicians may have never encountered the patient’s symptoms before, and gene-disease associations may not yet have been captured in the databases used by genome analysis software. Even comprehensive databases like ClinVar and gnomAD have gaps when it comes to rare disease variants. This makes diagnosing rare diseases extremely difficult. Shared data networks allow laboratories to pool anonymized patient data and apply matching algorithms to find unrelated individuals with similar phenotypes or genetic variations. This approach can help break down clinical and geographic silos and uncover variants of unknown significance, allowing laboratory technicians to identify patterns that would otherwise remain hidden and improve diagnostic potential.
“We are already seeing the benefits of a shared data network for patients,” Marshall says. “Even with just a few thousand samples, gnomAD This is extremely valuable for mutation evaluation. For example, if my patient has a rare mutation identified in another patient in the network, but the phenotype does not match, it can help rule it out. And if there is a match, it can help determine the diagnosis.” Data sharing has clear benefits, but it also raises important privacy challenges, such as information management, patient consent, and inter-institutional data agreements. Addressing these considerations and demonstrating clear clinical value is essential to expanding these networks.
Loureiro and Marshall believe that AI-assisted mutation analysis technology and shared data networks have the potential to provide more accurate and timely diagnoses, advance scientific research, accelerate the discovery of rare diseases, and ultimately facilitate diagnosis for rare disease patients and their families.
