Efforts in curation, data integration, and artificial intelligence are driving improvements in oncology variant interpretation, often a bottleneck between genomic data generation and actionable research insights, according to Illumina and Genomenon. The companies, who partnered to bring Genomenon’s Cancer Knowledgebase or “CKB” (formerly Clinical Knowledgebase) of curated variants to users of Illumina’s Connected Insights interpretation platform, say these efforts will enable scientists to interpret more oncogenic variants in more populations.
While rapid advances in genomics technologies have ushered in a new paradigm for precision medicine and made NGS more widely accessible — especially in oncology — new data, insights, and guidelines are evolving and expanding on a scale that poses challenges for the analysis and interpretation of results. This is contributing to what some call the variant interpretation bottleneck.
“The field is rapidly evolving,” said Cara Statz, a senior clinical analyst at Genomenon who began working on CKB nearly 10 years ago at The Jackson Laboratory, before it was acquired by Genomenon in the spring of this year. “Keeping up with all of the newest content coming out is a challenge.” On top of that, she said, the increasing availability and affordability of large NGS panels, exome sequencing, and genome sequencing mean that labs are finding more variants to interpret with each new type of assay.
“Modern NGS assays generate thousands and sometimes millions of variants from one case,” added Svetlana Bureeva, product manager at Illumina for Connected Insights, a clinical research platform enabling streamlined variant interpretation for oncology. Determining the relevance and actionability of variants on this scale significantly adds to test turnaround time. Bureeva noted a 2021 survey of specialists in advanced non-small cell lung cancer, in which 98 percent of respondents said they believe biomarker test results should be available within one or two weeks, but 37 percent said they wait an average of three to four weeks.
But with the efforts of Illumina, Genomenon, and other organizations, variant interpretation is gaining efficiencies, in part due to increasingly accessible technological innovations, said Statz and Bureeva. Genomenon provides two tools to facilitate timely variant interpretation: CKB and Mastermind. CKB is a database of manually curated information regarding actionability, treatment efficacy, and clinical trials related to over 45,000 cancer variants across 2,000 genes, enabling users to link molecular profiles to “report-ready” information on tumor type, treatment and trial options, and response prediction. Statz works with eight other PhD-level curators to not only add content to the knowledgebase, but also to maintain existing content, providing regular updates on approvals, guidelines, and trials in the US and internationally.
Mastermind is a tool with which users can search through variant data from journal articles, public databases, and Genomenon’s curated content. “CKB is a great source for providing answers for well-studied variants,” said Bureeva. “However, there are less-studied variants, and this is where Mastermind comes in.”
“But let's imagine there is nothing published about a variant. Nothing at all,” she added. “This is where computer predictors come in.” There are an estimated 70 million possible missense variants in the human genome. Though teams like CKB’s are working to understand and catalog the highest-impact pathogenic variants identified in the lab and the clinic, poorly understood or unstudied variants, known as variants of unknown significance (VUS), still pose a problem when profiling cancer or investigating genetic diseases.
“We are really only at the very beginning of our journey to understanding what variants do clinically in people,” said Kyle Farh, vice president and distinguished scientist at Illumina’s artificial intelligence lab, estimating that manually curating every missense variant found in humans will take thousands of years. “In the meantime, our team’s approach is to use deep learning to try to predict or impute the effects of variants based on what evidence we can find.”
In June 2023, Farh and colleagues published a study in Science describing PrimateAI-3D, an algorithm trained to estimate the potential pathogenicity of variants for research by referencing their frequency in other primate species. The algorithm takes advantage of the cross-species similarity of the protein-coding regions in primate genomes and the unlikelihood of natural selection allowing pathogenic variants to survive at high rates across species, Farh explained. While some groups have trained systems on human variant data from large databases, Farh and colleagues were concerned that the approach could result in weaknesses or biases in human interpretation being amplified by the algorithm. “So, instead of looking at what humans have called benign or pathogenic,” he said, “we looked at what evolution has called benign.” The algorithm includes analysis of 3D protein structures to predict the impact a missense variant could have on the function of the protein.
The resulting database of variants noted as benign or pathogenic is about 70 times larger than ClinVar, but Farh cautions that there are limitations to his team’s approach. “All you know is that evolution doesn't allow them to become frequent in the population,” he said. “So from that perspective, there's still a great deal of work that needs to be done, and it's not going to come just from comparative genomics. It's going to have to come from other clinical sources and functional experiments at high throughput.” Farh’s team had previously created SpliceAI, a neural network that predicts the pathogenicity of splicing variants. The prediction scores generated by PrimateAI-3D and SpliceAI are a core part of Illumina’s artificial intelligence algorithms found in their commercially available software, and available without charge for academic and non-profit research. Illumina’s AI capabilities are deployed within the Illumina Software ecosystem hosted on the Amazon Web Services (AWS) cloud to maximize the scalability and efficiency of computing algorithms.
Curation efforts like CKB, databases like Mastermind, predictive algorithms like PrimateAI-3D, and over 60 other sources of evidence for variant interpretation are integrated with Illumina’s Connected Insights, a unified platform for genomic analysis, interpretation, and reporting for clinical research applications. The platform is designed to manage and visualize genomic data, prioritize the top findings for potentially pathogenic variants, and determine meaningful insights based on the most up-to-date guidelines and other domain recommendations.
“What it means is that no finding is missed,” said Bureeva. Whereas labs may have previously referenced various databases and tools piecemeal, Connected Insights integrates each step of interpretation for clinical research. Now, more labs are seeing the value of integrating the workflow into one tool. “Because you don't only need to interpret the variants,” she said. “You need to have tools to upload data and put it all into a research report. You need to store that report. You need to have a tool to assess and share the report.” The result of the facilitated interpretation offered by the research platform, she said, is faster turnaround time, a more affordable comprehensive workflow, and more accessibility to different types of labs.
Genomenon and Illumina will be showcasing Genomenon’s Cancer Knowledgebase (CKB) and Illumina’s Connected Insights at the Association for Molecular Pathology (AMP) 2024 Annual Meeting on Nov. 19 through 23. Genomenon researchers will present two posters, one on using CKB to identify drug targets and the other on the creation of a database of somatic and germline RET variants created using CKB and Mastermind. Genomenon will also be presenting a corporate workshop on improving the interpretation of somatic and germline variants when working with whole genomes and large cancer panels (Wednesday, Nov. 20, 2:00 p.m. – 2:50 p.m.) and will be discussing its tools in booth 1233.
Farh’s team at Illumina will present research showing the performance of PrimateAI-3D and SpliceAI as compared to other classifiers, and Illumina will be giving live demonstrations of Connected Insights including the integrated knowledge bases, in booth 1107.
Learn more about Illumina Connected Insights here, and learn more about CKB here.