Abstract 2267: Assessment of AlphaMissense and structure-function predictions demonstrates efficient reclassification of genetic variants of unknown pathogenicity in inherited myeloid neoplasms

Abstract Introduction: Classification of genetic variants has significant implications for clinical management. Artificial Intelligence (AI) has the potential to transform classification of genetic variants, such as AlphaMissense, which incorporates structure-function relationships and allele frequencies across large datasets of genetic variants. While this represents a promising new tool, the performance of AlphaMissense has only been compared to known pathogenic and benign variants. Therefore, the discriminatory power for variants of unknown significance (VUS) has yet to be determined. We have performed a comparison of these predictions with variants associated with inherited myeloid neoplasms: 1) to assess the accuracy of these models against VUS; 2) to enhance classification of unpublished VUS by leveraging published information about other variants; and 3) to map the predicted classification onto protein structures to examine spatial patterns. Methods: For 83 genes associated with inherited myeloid diseases, a systematic literature review for missense variants was performed using the Mastermind Genomic Intelligence Platform and ClinVar. We used Alphamissense to characterize VUS missense variants according to predicted pathogenicity scores as pathogenic, ambiguous, or benign. Known pathogenic (n=1594) and benign (n=501) variants served as controls to evaluate the accuracy of the system. AlphaFold2 structures were visualized for structure-function rendering of reclassified VUS and known pathogenic variants to compare and identify discernible patterns of spatial distribution. Results: This variant dataset comprised 1594 pathogenic, 501 benign, and 46676 VUS missense variants. Among VUS, AlphaMissense reclassified a majority of VUS missense mutations and established that the system was 88% accurate in predicting pathogenicity. A higher percentage of variants was classified as pathogenic among tumor suppressor genes e. g. DDX41 and RUNX1 (57. 1% and 40. 5%) compared to the oncogenes, GATA1 and GATA2 (25. 7% and 34. 9%). Furthermore, 3D models revealed a majority of variants reclassified to pathogenic were located in regions with defined tertiary structure clustered with known pathogenic variants, while benign variants were located at peripheral positions lacking definite structure. Conclusions: These results demonstrate the discriminatory power of AlphaMissense for VUS pathogenicity prediction for both loss-of-function and gain-of-function disease mechanisms. This work may also have implications for rare or challenging somatic cancer variants and variants in rare diseases. Finally, incorporating information available from empirical and clinical studies for disease-causing variants offers the possibility of significantly enhancing the predictive power of these models. Citation Format: Amina Kurtovic-Kozaric, Asja Campara, Melissa Jahibasic, Amar Mujkic, Adnan Fojnica, Mark J. Kiel. Assessment of AlphaMissense and structure-function predictions demonstrates efficient reclassification of genetic variants of unknown pathogenicity in inherited myeloid neoplasms abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts) ; 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84 (6Suppl): Abstract nr 2267.