bio-vcf-statistics

What this skill does

## Version Compatibility

Reference examples tested with: bcftools 1.19+, numpy 1.26+

Before using code patterns, verify installed versions match. If versions differ:
- Python: `pip show <package>` then `help(module.function)` to check signatures
- CLI: `<tool> --version` then `<tool> --help` to confirm flags

If code throws ImportError, AttributeError, or TypeError, introspect the installed
package and adapt the example to match the actual API rather than retrying.

# VCF Statistics

Generate statistics and quality metrics using bcftools.

## Statistics Tools

| Command | Purpose |
|---------|---------|
| `bcftools stats` | Comprehensive variant statistics |
| `bcftools gtcheck` | Sample concordance and relatedness |
| `bcftools query` | Custom summaries |

## bcftools stats

**Goal:** Generate comprehensive variant statistics including counts, Ti/Tv ratio, and quality distributions.

**Approach:** Run bcftools stats and parse section-tagged output lines (SN, TSTV, AF, QUAL, DP).

**"How many variants are in this VCF?"** -> Compute summary counts, substitution types, and quality distributions from variant records.

### Basic Statistics

```bash
bcftools stats input.vcf.gz > stats.txt
```

### View Key Metrics

```bash
bcftools stats input.vcf.gz | grep "^SN"
```

Output sections:
- `SN` - Summary numbers
- `TSTV` - Transitions/transversions
- `SiS` - Singleton stats
- `AF` - Allele frequency distribution
- `QUAL` - Quality distribution
- `IDD` - Indel distribution
- `ST` - Substitution types
- `DP` - Depth distribution

### Summary Numbers (SN)

```bash
bcftools stats input.vcf.gz | grep "^SN" | cut -f3-
```

Reports:
- Number of samples
- Number of records
- Number of SNPs
- Number of indels
- Number of multiallelic sites
- Number of multiallelic SNPs

### Transition/Transversion Ratio

```bash
bcftools stats input.vcf.gz | grep "^TSTV"
```

Transitions (purine-to-purine: A↔G, or pyrimidine-to-pyrimidine: C↔T) are chemically favored over transversions because they preserve the purine/pyrimidine ring structure. CpG deamination (methylated C->T) is the single most common point mutation in vertebrate genomes and is a transition, which further inflates the Ti/Tv ratio. Exomes have higher Ti/Tv than whole genomes because coding regions are enriched for CpG dinucleotides.

Expected Ti/Tv ratio:
- Whole genome: ~2.0-2.1
- Exome: ~2.8-3.3
- Ti/Tv below expected range suggests excess false-positive SNPs (random errors produce Ti/Tv ~0.5)
- Ti/Tv above expected range suggests over-filtering that disproportionately removes transversions

### Per-Sample Statistics

```bash
bcftools stats -s - input.vcf.gz > per_sample.txt
```

### Compare Two VCFs

```bash
bcftools stats input1.vcf.gz input2.vcf.gz > comparison.txt
```

### Region-Specific Stats

```bash
bcftools stats -r chr1:1000000-2000000 input.vcf.gz > region_stats.txt
bcftools stats -R exome.bed input.vcf.gz > exome_stats.txt
```

## Plotting Statistics

**Goal:** Visualize variant statistics as publication-quality plots.

**Approach:** Pipe bcftools stats output to plot-vcfstats to generate PDF and PNG plots.

### Generate Plots

```bash
bcftools stats input.vcf.gz > stats.txt
plot-vcfstats -p output_dir stats.txt
```

Creates:
- `output_dir/summary.pdf`
- Individual PNG files

### Comparison Plots

```bash
bcftools stats file1.vcf.gz file2.vcf.gz > comparison.txt
plot-vcfstats -p comparison_dir comparison.txt
```

## Expected QC Metric Ranges

Interpreting variant statistics requires context-dependent thresholds. A metric that looks normal in WGS may be alarming in WES.

| Metric | WGS Expected | WES Expected | Flag If |
|--------|-------------|-------------|---------|
| Ti/Tv ratio | ~2.0-2.1 | ~2.8-3.3 | WGS: <1.8 or >2.5; WES: <2.5 or >3.5 |
| Het/Hom ratio | 1.5-2.0 | 1.5-2.0 | Outlier relative to cohort |
| Call rate per variant | >95% | >95% | <90% |
| Call rate per sample | >95% | >95% | <90% |
| Singleton rate (WGS) | ~100-200k | Variable | >2x cohort mean |
| Mendelian error rate (trios) | <0.5% | <0.5% | >1% |

Interpretation notes:
- Het/Hom ratio too high suggests contamination (mixed DNA inflates heterozygosity); too low suggests inbreeding, population structure, or excess homozygous reference calls from low coverage
- Excess singletons per sample suggest sample quality issues, contamination, or library artifacts
- Call rate thresholds apply to common variants; rare variants naturally have higher missingness
- These ranges assume human diploid germline calling; somatic, polyploid, or non-model organisms require different expectations

## Stratified Evaluation

A variant caller with 99% overall accuracy may perform at 70% in difficult genomic regions. Always evaluate stratified by region complexity.

```bash
bcftools stats -R easy_regions.bed input.vcf.gz > easy_stats.txt
bcftools stats -R difficult_regions.bed input.vcf.gz > difficult_stats.txt
```

Key stratification categories (using GIAB stratification BED files):
- **Easy/high-confidence regions** - baseline accuracy measurement
- **Homopolymer runs** - systematic indel errors, especially for Illumina and Ion Torrent
- **Tandem repeats / low-complexity** - alignment ambiguity inflates both FP and FN
- **Segmental duplications** - paralogous mapping produces false heterozygous calls
- **High GC (>70%) / Low GC (<25%)** - coverage bias creates systematic missingness
- **MHC / centromeric regions** - extreme polymorphism or repetitiveness defeats standard callers

GIAB stratification BED files are available at github.com/genome-in-a-bottle/genome-stratifications. Comparing Ti/Tv ratio between easy and difficult regions is a quick diagnostic: a drop in Ti/Tv within difficult regions confirms elevated false-positive rates there.

## Population-Scale QC Metrics

For multi-sample VCFs, per-sample and per-site population metrics reveal systematic issues invisible in single-sample analyses.

**Per-sample checks:**
- Variant count should be roughly consistent across samples from the same population (~4.5M SNPs per WGS sample in humans); outliers suggest processing errors or contamination
- Per-sample Het/Hom ratio: outliers suggest contamination (high) or sample swaps between populations (unexpected value)
- Excess singletons in one sample relative to the cohort mean suggests library or sequencing issues

**Per-site checks:**
- Excess heterozygosity (InbreedingCoeff < -0.3 in GATK, or HWE test): suggests genotyping error at sites where reads from paralogous regions pile up
- Hardy-Weinberg equilibrium: sites failing HWE (p < 1e-5) within a homogeneous population may be genotyping artifacts. HWE filtering must be applied within populations, never across mixed-ancestry cohorts where true allele frequency differences violate HWE assumptions

```bash
bcftools query -f '%CHROM\t%POS\t%INFO/InbreedingCoeff\n' input.vcf.gz | \
    awk '$3 < -0.3' > excess_het_sites.txt
```

## bcftools gtcheck

**Goal:** Verify sample identity and detect sample swaps by comparing genotype concordance.

**Approach:** Use bcftools gtcheck to compute pairwise discordance rates; interpret thresholds based on expected relatedness.

```bash
bcftools gtcheck -G 1 input.vcf.gz > relatedness.txt
bcftools gtcheck -g reference.vcf.gz query.vcf.gz
```

Discordance thresholds for interpreting results:
- Same individual (replicates/re-extractions): <0.5% discordance
- First-degree relatives (parent-child, siblings): ~10-15% discordance
- Unrelated individuals: >10% discordance (typically 15-25%)
- For trios: Mendelian error rate should be <0.5%; rates >1% suggest sample swaps or contamination

Output format (DC lines):

```
DC  0  sample1  sample2  0.95  1234  1200
```

Fields: DC tag, index, sample1, sample2, discordance rate, sites compared, discordant sites. Sample swaps are one of the most common errors in genomics studies. Running gtcheck is essential for any multi-sample study and should be performed early in the QC pipeline before downstream analysis.

## Quick Statistics