Abstract 6518: Integrating machine learning to optimize FFPE variant calling in a comprehensive genomic profiling assay for hematologic malignancies

Abstract Optimal clinical management of lymphoma and other hematological malignancies require assessment of somatic mutations across a subset of clinically relevant genes. However, most lymphoma specimens are received after being formalin fixed and paraffin-embedded (FFPE), a process that can induce DNA damage leading to reduced specificity when assayed with next generation sequencing (NGS). Here we present the analytical validation results of a targeted NGS panel of 141 clinically relevant genes for pan-heme indications using FFPE samples. We also describe an analysis pipeline that performs additional filtering steps for FFPE specimens. The analysis pipeline is designed to analyze FFPE and non-FFPE samples (blood/bone marrow samples from myeloid or lymphoid malignancies) sequenced on the same flow cell with specific analysis performed for each specimen type automatically. FFPE has distinct variant calling parameters compared to blood/bone marrow and an additional random forest machine learning (ML) model to filter variants associated with FFPE artifacts. The FFPE NGS panel interrogates all coding exons of the 141 genes to detect single nucleotide variants (SNVs) and insertions/deletions (indels) up to 50bp at variant allele frequency ≥ 5%. The custom hybrid capture-based assay utilizes genomic libraries created from 50-182.5 ng gDNA extracted from FFPE tissue, followed by sequencing on Illumina® instruments. Concordance studies were performed on clinical samples previously assessed using orthogonal NGS-based assays for SNVs/indels. In total, 198 FFPE samples including 72 unique clinical FFPE samples were assessed. Analysis of concordance demonstrated a positive percent agreement (PPA) of 93.9% for SNV/indels (388/413) and false discovery rate (FDR) of 4.9% (20/408). Assay precision was determined using three replicates of 5 clinical FFPE samples at minimal DNA input for both intra and inter-assay precision. Overall precision was 95.9% (394/411). Minimal DNA input was established to be 50ng of input material based on results obtained from 6 clinical FFPE samples. Analytical specificity was 99.99% for SNVs/indels based on 5 replicates of FFPE NA12878. Analytical sensitivity was 3.7% VAF for SNV/indels based on a hit rate dilution series. The ML model removed 10 putative FPs throughout the study. Performance on an independent test set of 44 clinical FFPE analysis showed concordance of 95.5% (445/466) compared to 89.5% (445/497) without the ML model. These data describe the FFPE performance of an assay that enables a comprehensive evaluation of genomic alterations in hematologic malignancies from all major specimen types and indications using one laboratory workflow. The workflow includes a flexible pipeline with an optional ML model to successfully filter artefacts associated with FFPE DNA damage. Citation Format: Grant Hogg, Tong Liu, Helen Cao, Adib Shafi, Amanda Williamson, Ashraf Shabaneh, Kimberly A. Holden, John Howitt, Xiaojun Guan, Michael Mooney, Li Cai, Eric A. Severson, Maria-Fernanda Senosain, Erik Vanroey, Shakti Ramkissoon, Anjen Chenn, Robert Daber, Marcia Eisenberg, Brian Caveney, Eyad Almasri, Taylor Jensen, Jon Williams. Integrating machine learning to optimize FFPE variant calling in a comprehensive genomic profiling assay for hematologic malignancies abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 6518.