Skip to main content
Premium Trial:

Request an Annual Quote

Genomic Medicine Calls for AI That Can Explain Itself, Illumina Says

By Illumina

While applications for artificial intelligence in genomic medicine and research abound, patients and clinicians are loath to trust often inscrutable black-box algorithms in their health decisions, where biases in models and training data create a direct impact on care. Artificial intelligence models that can explicitly demonstrate their logic and expose potential biases, known as explainable AI, are necessary for medical institutions to make healthcare decisions with increased confidence, according to Illumina. As part of its Connected Software ecosystem, Illumina offers Emedgene, a variant interpretation platform enabling evidence-backed insights and report generation for genetic disease and other germline applications.

Challenges arising from the ballooning volume of genetic data, the rising popularity of NGS-based tests, and the exponential increase in clinically relevant insights, together with the rarity of experts in genomic interpretation, are driving innovations in AI for genomic analysis, according to Niv Mizrahi, senior director of software development at Illumina, who is leading AI innovation for Emedgene. Curating information on the clinical relevance of gene variants, for example, is a time-consuming, laborious task that is not well suited for conventional software. Knowledge of genetic variants is scattered across publications and is generated from different study types, so curation has traditionally required human adaptability and comprehension. “This is where AI really comes into play,” said Mizrahi. “Can we curate all of this new knowledge that’s being learned from new studies, new clinical trials, new case studies, … into something cohesive that could actually be used in a lab-developed test environment?”

Genetic testing can benefit from AI at almost every stage, Mizrahi said. AI systems can extract and collate phenotypic data from health records, aid in test accessioning tasks, improve sequencing and secondary analysis, simplify variant classification and interpretation, and facilitate reporting. “A genetic test is many steps with many people doing many different things,” he said. “If you’re looking at the entire workflow of a physician meeting a patient to the physician getting the genetic test report from the lab, almost every part of the way could be leveraging machine learning or AI solutions in order to completely automate or at least enable geneticists and biologists to do that work.”

While machine learning algorithms are trained using reams of multidimensional data to make deep statistical associations, they are subject to the pitfalls of depending on spurious correlations from biased data, and they can make connections that humans don’t understand and can’t verify. “If you want your algorithms to be used, there has to be some element of trust,” Mizrahi said. “The goal of explainable AI [XAI] in clinical research is to say, ‘Okay, after we've got this prediction, what can I show a human being to comprehend and understand the prediction in the best way?’” XAI-powered systems can display the statistics involved in a software’s prioritized variant suggestions, additional information needed for a stronger prediction, potential changes in variables that would have the biggest impact on the prediction outcome, and decision trees visualizing the questions the XAI answered along the way.

Illumina’s Emedgene platform is a cloud-based, XAI-powered research tool that automates variant prioritization workflows for genetic and rare disease applications using whole-genome, whole-exome, or targeted panel data, including virtual panels. It presents the user with a small number of high-confidence candidate variants, typically five to 10, in order of priority for further review. Emedgene’s XAI feature, in conjunction with proprietary machine learning algorithms, visualizes the platform’s sources and calculations in a flow chart, illustrating connections between gene variants and phenotypes to allow for review and verification. It runs on top of a scalable, cloud-hosted environment using Amazon Web Services (AWS), enabling secure and rapid scaling of genomic data for customers. As part of Emedgene’s continuous cycle of releases, the feature was developed in 2019, long before "AI" became as popular as it is today. This technology suggests causative variants with high precision and speed, is explainable, and saves many hours of research. The platform was featured in a 2023 validation study done with Baylor Genetics and published in Genetics in Medicine in which the prioritized variants generated using XAI were validated against manual curation in 98 percent of subject trios, 93 percent of single subjects, and 97 percent overall for subjects analyzed.

This image is a screenshot from the Emedgene XAI tertiary analysis platform. It depicts a flow chart explaining the evidence and logic for a variant being suggested as causing a subject's condition. It identifies a heterozygous de novo variant in the PTPN11 gene known to cause Noonan syndrome. It identifies the subjects phenotypes of abnormal ventricular septum morphology, bilateral cryptorchidism, triangular face, and atrial septal defect as characteristic of Noonan syndrom, but lists other phenotypes present in the subject, such as abnormal bleeding, neoplasm, postnatal growth retardation, low posterior hairline, and intellectual disability as unconfirmed
Flow chart from Emedgene’s XAI interface explaining why a gene variant was prioritized as potentially causing Noonan syndrome in a subject.

Identifying variants in rare disease research in children on long diagnostic odysseys or in neonatal intensive care units is the “number-one use case” for the XAI functionality in Emedgene, said Mizrahi. Other research applications, which include other types of hereditary disease testing, carrier screening, pharmacogenomics, and healthy population-scale screening, lean heavily on Emedgene’s automated classification and curation capabilities. Additionally, pilot studies for newborn sequencing assay research are currently underway. Screening children for severe but treatable conditions is standard in many countries, but the number of conditions varies, and screening using metabolic tests may be limited in certain states or countries across the world, neglecting treatable pediatric rare diseases that could be detected with whole-genome sequencing, according to Mizrahi. “Enabling whole-genome newborn screening tests is my number one goal,” he said.

Mizrahi said he’s excited by the additional AI-enabled innovations being developed at Illumina, and that this is one strategic advantage of having this technology under one roof: as new AI tools are developed, they can be integrated into the Emedgene platform following the existing tenants of explainability, accuracy, and comprehensiveness. Two recent developments illustrate this advantage: the tight integration with DRAGEN Machine Learning (ML) and Multigenome (graph) enabled secondary analysis — proven for its accuracy, comprehensiveness, and efficiency — and the incorporation of the PrimateAI-3D missense prediction model. Primate AI-3D is a neural network for rare pathogenic variant detection trained on whole-genome data from 809 individual primates across 233 species and incorporating 3D protein structures. PrimateAI-3D, as described in Science by Kyle Kai-How Farh’s team at Illumina in June 2023, outperforms all known missense pathogenicity prediction methods. PrimateAI-3D algorithms, as part of Illumina’s latest Connected Annotations software, will soon be made available on secondary analysis variant callers including DRAGEN and Connected Analytics for a fully integrated NGS workflow. Primate AI-3D is integrated and available in Emedgene today.

In rare disease, where genetic test results can take weeks, and up to 60 percent of patients may not receive valuable results, the potential impact of these integrated AI algorithms and systems is immense, Mizrahi said. “We’re looking at how we can make workflows faster or simpler, and we believe AI models can advance clinical genomics and help improve patients’ lives.”

This sponsored content is provided by an advertiser and published in collaboration with the GW Custom Solutions Group, a division of GenomeWeb. The content was not produced by the editors or reporters of GenomeWeb, 360Dx, or Precision Oncology News, and does not represent the views of these publications or GenomeWeb's parent company, Crain Communications Inc.