Authors :
Presenting Author: Yi-Lee Ting, MS, CGC – Invitae
Swaroop Aradhya, PhD – Invitae; Alexandre Colavin, PhD – Invitae; Flavia Facio, MS, CGC – Invitae; Laure Fresard, MS, CGC – Invitae; Yuya Kobayashi, PhD – Invitae; Dianalee McKnight, PhD – Invitae; Keith Nykamp, PhD – Invitae; Jason Reuter, PhD – Invitae; Yi-Lee Ting, MS, CGC – Invitae; Britt Johnson, PhD – Invitae
Rationale: The clinical classification of missense variants is challenging due to limited evidence. Consequently, many remain categorized as variants of uncertain significance (VUS). VUS are at the core of healthcare disparities, as individuals from race, ethnicity and ancestry (REA) populations underrepresented in large genomic databases and medical literature tend to receive more VUS. To generate definitive genetic testing results more equitably across REA groups, we sought to leverage large genomic datasets by developing systematic machine learning (ML) models. Here, we evaluate the utility of these models in diagnostic multi-gene panel testing for epilepsy across REA groups.
Methods: From January 1, 2022 to May 22, 2023, gene-specific ML algorithms were validated and integrated at Invitae®. We incorporated existing models such as SpliceAI, and developed our own by leveraging large datasets including gnomAD, AlphaFold protein structures, and others, to model variant effects. Evidence from these ML models were incorporated into Sherloc, a semi-quantitative variant interpretation framework based on ACMG/AMP guidelines. Only evidence that met a negative or positive predictive value >80% was incorporated during variant interpretation. At least one validated model was available for 234 genes during the study period. VUS reduction was calculated and stratified by REA groups. Analyses of >20,000 patients were performed from random sampling of 20,000 patients with extrapolation. Measurement error was less than 2% variation by bootstrapping.
Results: Out of 57,668 US-based individuals that underwent diagnostic panel testing for epilepsy, ~55,100 (96%) had ML evidence applied to at least one variant and 25,393 (44%) were from an underrepresented group. Models contributed to classifying at least one benign/likely benign (B/LB) variant in ~36,100 (63%) and at least one pathogenic/likely pathogenic (P/LP) variant in ~1,900 (3%) individuals. Based on clinician-reported REA information at the time of requisition, a higher percentage of Asian (78%), Black (73%), and Hispanic (68%) individuals had an ML-dependent definitive classification (P/LP or B/LB) relative to White (59%) individuals by a one-tailed, two-sample proportion test (p < 3x10