一、Somatic vs Germline：本質不同
Germline calling 假設樣本是 diploid，每個位點 genotype 是 0/0、0/1、1/1 三選一，VAF 固定接近 0、0.5、或 1。但 tumor 樣本：
- 樣本不純：tumor purity 通常 30-90%，剩下是混入的 normal cells、stromal cells、infiltrating immune cells
- 克隆異質性：tumor 內部本身就是混合——early clonal mutations 出現在所有 tumor cells 中，subclonal mutations 只在部分
- VAF 是連續譜：可能 0.4（pure clonal het）、0.15（低純度 clonal）、0.05（subclonal）、甚至 0.01（深度測序看 ctDNA）
- copy number 可能異常：CNV 區域的 VAF 不再可預期
因此 somatic caller 不是 genotype caller，而是「在 noise 中找信號」的二元判斷：這個位點的 ALT 是真 mutation 還是 sequencing error / mapping artifact / contamination？
Germline calling assumes a diploid sample with three possible genotypes (0/0, 0/1, 1/1) and VAF fixed near 0, 0.5, or 1. But tumor samples:
- Impure samples: tumor purity is typically 30–90%, the rest is contaminating normal cells, stromal cells, infiltrating immune cells
- Clonal heterogeneity: tumors are internally mixed — early clonal mutations appear in all tumor cells, subclonal mutations only in some
- VAF is a continuum: could be 0.4 (pure clonal het), 0.15 (low-purity clonal), 0.05 (subclonal), even 0.01 (deep sequencing for ctDNA)
- Copy number may be abnormal: VAF in CNV regions is no longer predictable
So the somatic caller isn't a genotype caller — it's "signal hunting in noise" via binary judgment: is this site's ALT a real mutation, or a sequencing error / mapping artifact / contamination?
二、Mutect2：tumor-only vs tumor-vs-normal
Tumor-vs-Normal (推薦)
同病人取 normal tissue（通常 blood）作對照。任何 normal 中也存在的變異 = germline，可直接過濾掉，剩下的就是 somatic。最乾淨、最可信。
Use a normal tissue (usually blood) from the same patient as control. Any variant also present in normal = germline and is filtered out; what's left is somatic. Cleanest and most reliable.
Tumor-only
只有 tumor 樣本（normal 不可得，如 archived FFPE）。完全依賴 PoN + gnomAD AF 過濾 germline 與 artifacts。FP 率較高，需更嚴格 downstream filter。
Only tumor available (normal unobtainable, e.g., archived FFPE). Fully relies on PoN + gnomAD AF to filter germline and artifacts. Higher FP rate, stricter downstream filtering needed.
用 30+ 個健康樣本（同一定序中心、同 platform、同 capture kit），各自跑 Mutect2 tumor-only 模式，再用 CreateSomaticPanelOfNormals 合併成 VCF。任何「重複出現在多個 normal」的位點 = sequencing/alignment artifact，從 tumor 結果中過濾掉。PoN 是 Mutect2 tumor-only 的關鍵，缺乏 PoN 會被 artifact 淹沒。
Take 30+ healthy samples (same sequencing center, platform, and capture kit), run Mutect2 in tumor-only mode on each, then combine into a VCF with CreateSomaticPanelOfNormals. Any site repeatedly appearing across normals = sequencing/alignment artifact, filtered from tumor results. PoN is critical for Mutect2 tumor-only — without it, calls are drowned in artifacts.
三、互動模擬：tumor 內 VAF 分布
調整 tumor purity 與 subclonal fraction，看模擬的 1,000 個 somatic mutations 在 VAF 直方圖上的分布。Clonal mutations 的中位 VAF ≈ purity/2 (heterozygous in all tumor cells)；subclonal mutations 形成 VAF 較低的尾巴。
Adjust tumor purity and subclonal fraction to see how 1,000 simulated somatic mutations distribute in the VAF histogram. Clonal mutations have median VAF ≈ purity/2 (heterozygous in all tumor cells); subclonal mutations form a low-VAF tail.
四、Mutect2 完整 pipeline
| 步驟 | 工具 | 目的 |
|---|---|---|
| 1 | Mutect2 | 呼叫候選 somatic variants (tumor + normal + PoN + germline AF) |
| 2 | GetPileupSummaries | 計算 tumor 與 normal 在 common variants 的 allele counts |
| 3 | CalculateContamination | 估計 cross-sample contamination 比例 |
| 4 | LearnReadOrientationModel | 學 FFPE oxidation / 8-oxoG artifact 的 read orientation bias |
| 5 | FilterMutectCalls | 整合上述所有資訊，標記 PASS 或具體 filter 原因 |
| 6 (選) | FilterAlignmentArtifacts | 重新對齊到 BWA-MEM2 + alt aware 參考，剔除 mapping artifacts |
FFPE (formalin-fixed paraffin-embedded) 樣本是病理檔案常見來源，但 formalin 會氧化 G→8-oxoG，定序時讀為 T，產生系統性 G>T / C>A artifact。LearnReadOrientationModel 會學出此 bias 的 strand-specific pattern，FilterMutectCalls 用 --ob-priors 套用，可大幅減少 FP。新鮮血液 / 凍存組織不必擔心，但 FFPE 一定要做這步。
FFPE (formalin-fixed paraffin-embedded) samples are common from pathology archives, but formalin oxidizes G→8-oxoG, read as T during sequencing — producing systematic G>T / C>A artifacts. LearnReadOrientationModel learns this bias's strand-specific pattern; FilterMutectCalls applies it via --ob-priors, drastically reducing FPs. Fresh blood / frozen tissue doesn't need this, but FFPE absolutely does.
五、Pipeline 指令範例
# Mutect2 — tumor + matched normal
gatk Mutect2 \
-R Homo_sapiens_assembly38.fasta \
-I tumor.bam \
-I normal.bam \
-tumor TUMOR_SAMPLE_ID \
-normal NORMAL_SAMPLE_ID \
--germline-resource af-only-gnomad.hg38.vcf.gz \
--panel-of-normals 1000g_pon.hg38.vcf.gz \
--f1r2-tar-gz tumor.f1r2.tar.gz \
-O tumor.unfiltered.vcf.gz
# 重要：tumor 樣本名要與 BAM @RG SM tag 完全一致# Mutect2 — tumor-only mode
gatk Mutect2 \
-R Homo_sapiens_assembly38.fasta \
-I tumor.bam \
-tumor TUMOR_SAMPLE_ID \
--germline-resource af-only-gnomad.hg38.vcf.gz \
--panel-of-normals 1000g_pon.hg38.vcf.gz \
--f1r2-tar-gz tumor.f1r2.tar.gz \
-O tumor.unfiltered.vcf.gz
# 注意：tumor-only 模式對 PoN 高度依賴；若無自家 PoN，
# Broad 提供 1000G PoN 可下載 (gs://gatk-best-practices/somatic-hg38/)# 建立自家 PoN (建議用 30+ 個與 tumor 同 platform/capture 的 normals)
# 1. 每個 normal 跑 Mutect2 tumor-only 模式
for normal in normal_001 normal_002 ... normal_030; do
gatk Mutect2 -R ref.fa -I ${normal}.bam -tumor ${normal} \
--max-mnp-distance 0 \
-O pon/${normal}.vcf.gz
done
# 2. GenomicsDBImport
gatk GenomicsDBImport \
-R ref.fa -L wgs_calling_regions.hg38.interval_list \
--genomicsdb-workspace-path pon_db \
$(for v in pon/*.vcf.gz; do echo -n "-V $v "; done)
# 3. 合併為 PoN VCF
gatk CreateSomaticPanelOfNormals \
-R ref.fa \
--germline-resource af-only-gnomad.hg38.vcf.gz \
-V gendb://pon_db \
-O my_panel_of_normals.vcf.gz# Step 1: contamination
gatk GetPileupSummaries -I tumor.bam \
-V small_exac_common_3.hg38.vcf.gz \
-L small_exac_common_3.hg38.vcf.gz \
-O tumor.pileups.table
gatk GetPileupSummaries -I normal.bam \
-V small_exac_common_3.hg38.vcf.gz \
-L small_exac_common_3.hg38.vcf.gz \
-O normal.pileups.table
gatk CalculateContamination \
-I tumor.pileups.table \
-matched normal.pileups.table \
-O tumor.contamination.table \
--tumor-segmentation tumor.segments.table
# Step 2: orientation bias model (if FFPE)
gatk LearnReadOrientationModel \
-I tumor.f1r2.tar.gz \
-O tumor.read-orientation-model.tar.gz
# Step 3: 整合 filter
gatk FilterMutectCalls \
-R ref.fa \
-V tumor.unfiltered.vcf.gz \
--contamination-table tumor.contamination.table \
--tumor-segmentation tumor.segments.table \
--ob-priors tumor.read-orientation-model.tar.gz \
-O tumor.filtered.vcf.gz
# 只留 PASS
bcftools view -f PASS tumor.filtered.vcf.gz -Oz -o tumor.somatic.vcf.gz六、TMB 與下游應用
過濾後的 somatic VCF 是癌症基因體分析的起點，常見下游：
- TMB (Tumor Mutation Burden)：每 Mb 非同義 somatic mutations 數。TMB-high (typically ≥ 10 mut/Mb) 與 immunotherapy (anti-PD-1) 反應正相關，已是 FDA-approved biomarker。
- Mutational signatures (COSMIC SBS)：deconvolute 96 trinucleotide context 計數，找出底層 mutagenesis (UV、tobacco、APOBEC、MMR-deficient)。SBS4 = tobacco, SBS7 = UV。
- Driver gene identification：對照 COSMIC Cancer Gene Census、OncoKB；用 dN/dS、MutSig、IntOGen 識別驅動突變。
- Actionability annotation：對照 OncoKB、CIViC、JAX-CKB 找 FDA-approved drug targets，分 AMP Tier I-IV。
The filtered somatic VCF is the starting point for cancer genomic analysis. Common downstream uses:
- TMB (Tumor Mutation Burden): non-synonymous somatic mutations per Mb. High TMB (typically ≥10 mut/Mb) correlates with immunotherapy (anti-PD-1) response — an FDA-approved biomarker.
- Mutational signatures (COSMIC SBS): deconvolute 96-trinucleotide-context counts to identify underlying mutagenesis (UV, tobacco, APOBEC, MMR-deficient). SBS4 = tobacco, SBS7 = UV.
- Driver gene identification: cross-reference COSMIC Cancer Gene Census, OncoKB; use dN/dS, MutSig, IntOGen to identify driver mutations.
- Actionability annotation: cross-reference OncoKB, CIViC, JAX-CKB to find FDA-approved drug targets, categorized by AMP Tier I–IV.
七、FAQ
tumor coverage 該打多少？
contamination 多少算嚴重？
Mutect2 與 Strelka2 / VarDict / DeepVariant 比較？
如果沒有 matched normal 怎麼辦？
八、小測驗
Q1. 為什麼 somatic caller 不是 genotype caller？
Q2. PoN 主要過濾掉的是？
Q3. FFPE 樣本的特徵性 artifact 是？