Illumina COLO829 tumour-only, 50x, somatic SNVs (Chen et al. 2025 Supplementary Table 7b)

Stable ID: somatic-clairs-to-20261011-protocol-illumina-colo829-snv

areas

dna-genomes

contexts

clinical_research

protocol

Each caller is run on the tumour BAM alone; calls are compared with the filtered cell-line truth set using som.py (hap.py 0.3.12) and ClairS-TO compare_vcf, which the authors cross-checked. Precision, recall and F1 are reported at the quality threshold giving the best F1.

version

Supplementary Table 7b, Type 'SNV'

metric

f1-score

limitations

Developer paper: the authors developed ClairS-TO and its SS and SSRS models (author_reported); Clair3, also run here, comes from the same group. Comparators were run by the authors with their recommended settings.; Tumour-only calling of cell lines; the truth sets were derived from tumour-normal data and are filtered to variants with coverage at least 4, at least 3 alternative reads and VAF at least 0.05, so very low-VAF variants are not scored.; Precision, recall and F1 are at the best achievable F1 over the caller's quality threshold, chosen on the test data (Figure 2 legend); AUPRC summarises the whole curve. Neither is an operating point fixed in advance.; Clair3 is a germline caller; its calls were treated as somatic after removing those tagged by ClairS-TO's panels of normals.; Single sample per dataset; no interval estimates.; Comparator versions are old relative to the paper (DeepSomatic 1.7.0 run from the 'latest' container; Mutect2 4.2.6.1; Octopus 0.7.4; Pisces 5.3.0).; Competing interests: the senior author (R.L.) receives research funding from ONT.; Short-read comparators (DeepSomatic WGS_TUMOR_ONLY, Mutect2 with gnomAD germline resource, Octopus, Pisces) were run in tumour-only mode; the panel-of-normals and filtering choices for each are in the Supplementary Information command lines.; TP + FN is not identical across callers (SNV 40,232, but 40,230 for Pisces; indel 715, but 714 for Mutect2 and 713 for Octopus), so the scored truth set differs slightly between rows; the source does not explain this.; Whether the DeepSomatic 1.7.0 models were trained with COLO829 data is not stated in the source.

source locator

Supplementary Table 7b; Results 'Benchmarking datasets and evaluation'; Methods 'Benchmarking'; Supplementary Information 'Command lines used'

• uses data: COLO829 tumour Illumina NovaSeq 6000 WGS (ClairS-TO benchmark)
• assessment: ClairS-TO SS on illumina-colo829 tumour-only SNV (Chen et al. 2025)
• assessment: ClairS-TO SSRS on illumina-colo829 tumour-only SNV (Chen et al. 2025)
• assessment: DeepSomatic on illumina-colo829 tumour-only SNV (Chen et al. 2025)
• assessment: Mutect2 on illumina-colo829 tumour-only SNV (Chen et al. 2025)
• assessment: Octopus on illumina-colo829 tumour-only SNV (Chen et al. 2025)
• assessment: Pisces on illumina-colo829 tumour-only SNV (Chen et al. 2025)
• assessed by: Select a tumour DNA somatic variant-calling workflow