AlphaMissense performance as a classifier of germline and somatic variants of unknown significance in high-grade sarcoma.

11538 Background: Missense variants associated with high frequency, clinical context, splicing, penetrance and biochemical function are scored as pathogenic by ClinVar/ACMG (scores >99%/>90%). Pathogenic scores that range from this cut-point to 10% (benign) are referred to as variants of unknown significance (VUS). Alphafold2 is a deep neural network algorithm for protein structure trained on known molecular structures. It has informed AlphaMissense (AM) where the impact of every possible single missense variant in the human genome is assessed. The AM continuous pathogenicity score (roughly 0-0.35 benign, 0.36-0.55 ambiguous, 0.56-1.0 pathogenic) are claimed to have 90% accuracy based on ClinVar (based on 1,263 pathogenic and 1,263 benign variants). Methods: Here, we identified VUSs from the Oxford Sarcoma Precision Oncology real world cohort. This included the identification of somatic variants from FoundationOne Heme Sarcoma analysis of Formalin-Fixed Paraffin Embedded (FFPE) tumour samples. Germline variants were identified from whole exome sequencing (WES) (97 genes) from patient blood samples followed by analysis using GATK guidelines. All VUS identified were further analysed using AM. Results: AM reclassified the identified >1000 VUSs from 60 high grade sarcoma gene panels predominantly into benign or pathogenic, with only ~1% remaining as ambiguous. Variants reclassified as pathogenic were assessed for likelihood of pathogenicity based on scientific rationale. This involved evaluating current literature on the variant, determining evolutionary consensus of the residue, structural analysis to assess the potential function of the residue and enrichment of the variant within sarcoma patients compared to known frequencies across the population. Conclusions: In conclusion, the AM classifier dichotomises VUS to either benign or pathogenic, but this is not based on either structural or functional predictions directly. We conclude that it should not be currently used in the analysis of clinical samples seeking actionable variants but could instead stratify potential pathogenic variants for functional investigation.