For patients and families affected by rare diseases, the path to a genetic diagnosis can be long, fragmented, and emotionally exhausting. On average, it takes five years from symptom onset to a definitive molecular diagnosis,¹ and for some groups, including adults with rare diseases, the journey may extend to two decades.² Exome and genome sequencing have transformed our ability to identify potential genetic contributors to disease, but generating sequence data is only the beginning. The greater challenge is interpretation: determining which variants may explain a patient's clinical presentation and which are unlikely to be relevant.

Although standards exist for variant classification,³ the process remains labor intensive, and many variants lack sufficient evidence to support classification as pathogenic/likely pathogenic or benign/likely benign. As a result, they are classified as variants of uncertain significance (VUS). In a study of 1.6 million individuals undergoing genetic testing, approximately 40% had at least one VUS.⁴ This burden is reflected in ClinVar, where approximately 2.3 million variants were classified as VUS as of August 31, 2026.⁵ Despite predictions from the National Human Genome Research Institute that VUS would largely disappear by 2030,⁶ the number of VUS reported in ClinVar continues to increase. Moreover, VUS rates are higher among individuals from racial and ethnic groups other than European White populations,⁴ contributing to persistent inequities in genomic medicine.

Artificial intelligence (AI)-informed approaches have been proposed as a way to improve variant interpretation and shorten the diagnostic odyssey,⁷ but can they reduce uncertainty without amplifying existing disparities? Emerging evidence suggests they may. A tool combining large language models and Bayesian inference reclassified more than 1,000 VUS across 17 genes, affecting approximately 45,000 patients who had previously received a VUS result.⁸ Notably, it achieved this using the same clinical evidence available to laboratory analysts. Another study demonstrated that machine learning models integrating biological data and variant frequency information reduced VUS rates more effectively among historically underrepresented populations.⁹

These findings highlight an important point: machine learning models can integrate and interpret vast amounts of information in ways that exceed human cognitive capacity.¹⁰ However, their performance depends on the quality and diversity of training data and the expert knowledge embedded within their design. The advances reported thus far have been achieved through collaboration between AI and human experts, not AI alone. For genetics professionals, developing fluency in AI is increasingly important if we are to help shape the future of genomic medicine.¹¹

In our laboratory, variant interpretation combines genomic data, clinician-provided phenotypic information, AI-informed decision-support models, and expert review. Central to this process is our internal rare disease genomic knowledgebase—GeneDx Infinity™. It integrates data from more than one million sequenced exomes and genomes, including gene-disease associations, variant classifications, internal frequency data, clinical phenotypes, and familial testing information accumulated over 25 years of rare disease testing. This breadth matters because genomic interpretation is strongest when sequence data are evaluated in clinical context, informed by public databases, published literature, and real-world clinical experience.

Infinity captures the diversity of rare disease, with approximately half of the dataset representing individuals of non-European ancestry,¹² and spans a broad range of phenotypes, including autism, congenital heart defects, cerebral palsy, epilepsy, and multiple congenital anomalies among others. Although representation alone cannot eliminate disparities, more diverse datasets help reduce the likelihood that AI models are optimized for only a subset of patients. For genetic counselors, this reinforces a fundamental principle of equitable genomic medicine: more representative data can improve interpretation and reduce uncertainty for underrepresented populations.

The value of Infinity lies not only in its scale but also in its clinical depth. Variants become more informative when interpreted alongside phenotype, inheritance patterns, prior testing, and evidence accumulated across similar cases. Clinical notes, Human Phenotype Ontology terms, and parental samples provide critical context, with more than 60% of GeneDx exome and genome cases including parental data. These resources enable analysts to move beyond asking, “What variants are present?” to the more clinically meaningful question, “Which finding, if any, best explains this patient's presentation?”

AI-informed models are essential for leveraging these large-scale data resources in an efficient and scalable manner.  They are not diagnostic replacements or autonomous decision-makers. Rather, they are tools that help clinical experts focus their attention on the findings most likely to warrant deeper review. In our laboratory these models are informed and refined by board-certified laboratory geneticists and genetic counselors.

One example is Multiscore, an AI-powered gene prioritization model integrated into the GeneDx interpretation framework.¹³ Multiscore evaluates how well each gene's known phenotype associations align with a patient's clinical presentation using information from multiple sources of gene–phenotype knowledge, including OMIM, published cases, and internal clinical data. In a published evaluation of 9,989 exome and genome cases, Multiscore ranked the disease-causing gene among the top ten candidates in 84% of cases. While prioritization does not establish a molecular diagnosis, it reduces the number of genes requiring review, enabling experts to focus on nuanced interpretation while retaining flexibility to evaluate genes outside the prioritized list when clinically warranted.

Another AI-informed tool, MIMI Ranker, extends this approach to variant-level prioritization by integrating variant-level evidence, phenotype information, allele frequency data, and gene-level knowledge from Infinity.¹⁴ Across 16,000 clinical exome and genome cases, MIMI Ranker placed all pathogenic and likely pathogenic variants within the top three candidate variants in approximately 98% of diagnostic cases. Notably, its performance improves as additional diagnostic cases are incorporated into the training dataset. Together, Multiscore and MIMI Ranker help narrow a vast search space so that expert reviewers can focus attention where it is most likely to have clinical impact.

For genetic counselors, AI-informed tools have the potential to shorten diagnostic odysseys, reduce the burden of VUS, and support clearer diagnostic pathways. Yet, as emphasized in the NSGC position statement on The Use of Artificial Intelligence in Genetic Counseling,¹⁵ AI is most powerful when paired with human expertise. While AI can prioritize data at a scale beyond human capability, it cannot replace clinical judgment, contextual interpretation, or patient-centered care. The future of genomic diagnosis will be shaped not by AI alone, but by the thoughtful integration of high-quality data, expert oversight, and compassionate care to transform genomic information into answers for patients and families.

References

1. Faye, F., Crocione, C., Anido de Peña, R., Bellagambi, S., Escati Peñaloza, L., Hunter, A., Jensen, L., Oosterwijk, C., Schoeters, E., de Vicente, D., Faivre, L., Wilbur, M., Le Cam, Y., & Dubief, J. (2024). Time to diagnosis and determinants of diagnostic delays of people living with a rare disease: results of a Rare Barometer retrospective patient survey. European Journal of Human Genetics: EJHG, 32(9), 1116–1126. https://doi.org/10.1038/s41431-024-01604-z 2. Schuermans, N., Hemelsoet, D., Terryn, W., Steyaert, S., Van Coster, R., Coucke, P. J., Steyaert, W., Callewaert, B., Bogaert, E., Verloo, P., Vanlander, A. V., Debackere, E., Ghijsels, J., LeBlanc, P., Verdin, H., Naesens, L., Haerynck, F., Callens, S., Dermaut, B., Poppe, B., … for UD-PrOZA (2022). Shortcutting the diagnostic odyssey: the multidisciplinary Program for Undiagnosed Rare Diseases in adults (UD-PrOZA). Orphanet Journal of RareDiseases, 17(1), 210. https://doi.org/10.1186/s13023-022-02365-y 3. Richards, S., Aziz, N., Bale, S., Bick, D., Das, S., Gastier-Foster, J., Grody, W. W., Hegde, M., Lyon, E., Spector, E., Voelkerding, K., Rehm, H. L., & ACMG Laboratory Quality Assurance Committee (2015). Standards and guidelines for the interpretation of sequence variants: a joint consensus recommendation of the American College of Medical Genetics and Genomics and the Association for Molecular Pathology. Genetics in Medicine: official journal of the American College of Medical Genetics, 17(5), 405–424. https://doi.org/10.1038/gim.2015.30 4. Chen, E., Facio, F. M., Aradhya, K. W., Rojahn, S., Hatchell, K. E., Aguilar, S., Ouyang, K., Saitta, S., Hanson-Kwan, A. K., Capurro, N. N., Takamine, E., Jamuar, S. S., McKnight, D., Johnson, B., & Aradhya, S. (2023). Rates and Classification of Variants of Uncertain Significance in Hereditary Disease Genetic Testing. JAMA network open, 6(10), e2339571. https://doi.org/10.1001/jamanetworkopen.2023.39571 5. Landrum MJ, Lee JM, Riley GR, Jang W, Rubinstein WS, Church DM, Maglott DR. ClinVar: public archive of relationships among sequence variation and human phenotype. Nucleic Acids Res. 2014 Jan 1;42(1):D980-5. doi: 10.1093/nar/gkt1113. [PubMed PMID: 24234437] 6. Gunter, C., & Green, E. D. (2023). To boldly go: Unpacking the NHGRI's bold predictions for human genomics by 2030. American Journal of Human Genetics, 110(11), 1829–1831. https://doi.org/10.1016/j.ajhg.2023.09.010 7. Wilk, E. J., Taluri, S., Howton, T. C., Crumley, A. B.,Mrug, M., & Lasseigne, B. N. (2026). AI in variant analysis: fast track to genetic diagnoses. Human Genetics, 145(1), 54. https://doi.org/10.1007/s00439-026-02847-0 8. Manders, T. R., Tan, C. A., Kobayashi, Y., Wahl, A., Araya, C., Colavin, A., Facio, F. M., Metz, H., Reuter, J., Frésard, L., Padigepati, S. R., Stafford, D. A., Nussbaum, R. L., & Nykamp, K. (2025). Harnessing genotype and phenotype data for population-scale variant classification using large language models and bayesian inference. Human Genetics, 144(6), 605–614. https://doi.org/10.1007/s00439-025-02743-z 9.Fresard, L., Facio, F.M., Chen, E., Colavin, A., Johnson, B., Araya, C., Manders, T., Wahl, A., Metz, H., Nicoludis, J., Ouyang, K., Padigepati, S., Kobayashi, Y., Reuter, J., Nykamp, K. (2025) The impact of systematized generation, evaluation, and incorporation of machine learning algorithms for clinical variant classification. MedRxiv 2025.02.03.25321356; doi: https://doi.org/10.1101/2025.02.03.25321356  10. Aradhya, S., Facio, F. M., Metz, H., Manders, T., Colavin, A., Kobayashi, Y., Nykamp, K., Johnson, B., & Nussbaum, R. L. (2023). Applications of artificial intelligence in clinical laboratory genomics. American Journal of Medical Genetics. Part C, Seminars in medical genetics, 193(3), e32057. https://doi.org/10.1002/ajmg.c.32057 11. Solomon B. D. (2026). Artificial intelligence in genomic medicine: dispelling three myths. NPJ Genomic Medicine, 11(1), 24. https://doi.org/10.1038/s41525-026-00575-y 12. Kelly K, Napier MP, Poll S, Mitchell D, King K, Cruickshank J, Johnson B, Facio FM. (March 2026) The Diagnostic Yield of Exome and Genome Sequencing in a Large and Diverse Cohort of Over 300,000 Patients: Context Matters, Presented at ACMG Annual Clinical Genetics Meeting, Baltimore, MD.​ 13. Ustach, V. D., Guillen Sacoto, M. J., McGee, S., Gainullin, V. G., Arvai, K., Begtrup, A., Facio, F. M., Greenberg, M., Guðbjartsson, H., McWalter, K., Millán, F., Monaghan, K., Retterer, K., Richard, G., Topaz, N., Torene, R., Johnson, B., & Laurent, T. (2026). Multiscore, a gene ranker powered by artificial intelligence and real-world clinical data, shows high sensitivity for the molecular diagnosis of Mendelian disorders in nearly 10,000 exomes and genomes. Human genetics, 145(1), 22. https://doi.org/10.1007/s00439-025-02797-z 14. Begtrup A, Guðbjartsson H, Guillen Sacoto M, Olafsdottir H, Greenberg M, Facio FM, Ustach VD. (March 2026) MIMI Ranker: An AI-Powered Tool for Pathogenic Variant Prioritization Tested on a Diverse Dataset of 16K Clinical Exomes and Genomes, Presented at ACMG Annual Clinical Genetics Meeting, Baltimore, MD.​ 15. Adopted 2026

Flavia Facio, MS, CGC

is Medical Affairs Director, Platform Technologies at GeneDx. She is a thought leader in genomic research, with a proven track record of generating the evidence needed to realize the promise of personalized medicine. Over her 25-year career spanning clinical, academic, and industry settings, she has led strategic research programs and teams focused on translating emerging technologies into clinical practice. As a clinician-scientist, she brings a unique perspective on the needs of patients, the realities of clinical care, and the gaps in the evidence base for genomic medicine.