Abstract Artifactual variants obfuscate the already difficult task of somatic variant calling by mimicking some qualities of low variant allele fraction (VAF) mutations in this low signal-to-noise ratio regime. Statistical callers cannot cover the entire landscape of artifacts, leaving a critical gap. While deep learning tools for variant calling exist, they generally consume enormous amounts of labeled training data, are expensive to run, and can be overfit to a single sequencing technology. Here we present and benchmark Permutect, a lightweight deep learning-based variant caller for identifying artifactual variants across sequencing technologies and genomes. Permutect combines an artifact model for technological context and posterior model for biological context, respectively, allowing it to learn the characteristics not only for the technology but additionally the particularities of a given lab environment. The tool leverages the inherent non-orderedness of a set of reads (and thus their permutation invariance) to ascribe a probability for being a technical artifact. Permutect is therefore a lightweight deep learning model which can use as little as a single (unlabeled) genome for training thereby addressing the shortcomings of existing statistical and deep-learning callers. We tested this novel tool by training on four Genome-in-a-Bottle samples and benchmarking it against Mutect2, Strelka2, and DeepSomatic across well-established callsets: the ICGC Dream Challenge Sets 1-4 and SEquencing Quality Control Phase 2 (SEQC2) HiSeq and NovaSeq replicates. Permutect is an effective and precise tool for filtering, particularly in the paired Tumor-Normal (TN) setting. It achieved the highest mean Precision score (0.928) and the highest mean F1 score (0.899) in the Dream TN cohort, demonstrating superior overall balanced performance relative to its traditional competitors. Performance remained high and consistent across SeqC2 data (e.g., F1 score of 0.941 on HiSeq TN); we exclude scores from DeepSomatic, which was trained on that same data. While the precision of all tools are depressed in the Tumor-Only (TO) setting, it maintains a strong mean Recall (0.810) comparable to that of Mutect2 (0.811) in Dream TO. Though all callers struggled in the TO context, our findings show that Permutect's performance is competitive, and often superior, to established methods, highlighting its potential to address the pervasive issue of artifactual variants. Further work is being done to extend the domain of Permutect to whole-exome and long read sequencing and other improvements such as including the tri-nucleotide context about a variant will further refine the tool to address the shortcomings seen in the TO setting. In doing so, we expect Permutect to enhance downstream genomic analysis with applications ranging from clinical molecular pathology to the design of personalized immunotherapies. Citation Format: Julian Gascoyne, David Benjamin, Juan Gallegos, Lee T. Lichtenstein, Sachet Shukla. A novel, deep learning -based filtering tool for enhanced detection of technical artifacts in whole-genome sequencing data 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 1495.
Gascoyne et al. (Fri,) studied this question.