SomatiX: accurate and efficient somatic SNV detection from matched tumor-normal long-read DNA sequencing.

SomatiX achieves mean F1 scores of 0.986 for ONT and 0.988 for PacBio on held-out chromosome 1 cancer cell-line benchmarks at ≥3 variant-supporting reads and VAF ≥0.05.

• Why it matters: Accurate somatic SNV calling from matched tumor-normal long-read data is hindered by sequencing errors, variable tumor purity, tumor-in-normal contamination, and low VAFs. Robust low-VAF detection and efficient genome-scale analysis remain key challenges despite prior deep-learning progress.
• What they did: SomatiX is a deep-learning somatic SNV caller built on an allele-aware pileup representation, benchmarked across diverse cancer cell-line datasets, VAF bins, sequencing contexts, and read-support thresholds. It was compared against DeepSomatic and ClairS using F1 score across ONT and PacBio conditions.
• The result: In comprehensive benchmarks, SomatiX outperformed DeepSomatic and ClairS in F1 score across most ONT benchmark conditions and showed competitive performance on PacBio data. It particularly performed strongly for low-VAF SNVs in ONT while using a compact model and achieving the fastest runtime among evaluated tools.