Performance and validation of a tumor mutation profiling, based on artificial intelligence annotation, to assist oncology decision making.

Abstract

e13148 Background: Tumor mutation profiling has become a key component for orienteering the treatment of oncologic patients. A crucial step for this is the correct identification and classification of pathogenic and actionable variants. In the present work we aimed at the development and validation of a tumor mutation profiling panel, based on NGS, which uses artificial intelligence for variant annotation. Methods: We designed a hybrid capture panel, containing 366 genes to evaluate somatic SNVs, INDELs and CNVs, and to calculate TMB, using a customized bioinformatics pipeline. MSI status was determined by fragment analysis using capillary electrophoresis. Analytical performance was determined using reference cell lines. FFPE samples from 70 tumors were accessed and 53 were sequenced. Variant annotation was performed by IBM Watson for Genomics (WfG) platform. Assay performance on clinical samples was defined based on orthogonal assays using Agilent CGH+SNParray 400K (for CNVs only) and Foundation One test (Foundation Medicine) (F1). Results: Breast, colon and lung were the most common tumor origins. Fifty-three samples were successfully sequenced, while 41 of them could also be analyzed by F1 test. A summary of the assay performance is presented in Table 1. Our pipeline detected 1219 variants and 290 (23%) were classified as Pathogenic, Likely Pathogenic or Actionable, according to WfG. Thirty-five samples (66%) presented a variant that could drive the treatment, with 37.7% of samples being sensitive to targeted therapies, while 22.6% were resistant; additionally, 86% had an indication for a clinical trial. Conclusions: The developed assay presented a good overall sensitivity and allele frequency correlation, with TMB and MSI having the best rates. Comparisons with F1 had reduced values of concordance; however, SNVs and INDELs presented a similar frequency. Differences on CNVs identification may rely on distinct thresholds established by the different groups. The high percentage of samples that could benefit from mutational profiling highlights the importance of such approach in the clinical routine. Additionally, the high number of variants features the need for updated information for annotation. [Table: see text]