# Structural variants (Cancer GIAB)

**TL;DR:** the Cancer Genome in a Bottle project publishes HG008, a matched
tumor/normal PDAC cell line, as PacBio HiFi reads, a draft SV and CNV benchmark,
and a telomere-to-telomere assembly of the tumor. Load them as tracks and read
each benchmark call against the alignments and the copy number under it. The
tumor is hypodiploid, so arm-level loss is the backdrop to every call.

## Prerequisites

The walkthroughs at the end need none of this: they run on the finished site,
hosted at
[the C-GIAB demo](https://jbrowse.org/code/jb2/main/?config=https://jbrowse.org/demos/cgiab/config.json).
Building your own instance needs:

- A machine with HTTP access, either a public URL or `http://localhost`
- ~1 TB of free disk to build the tracks from the raw reads, or ~1.5 TB for the
  [full pipeline](#reproduce-it-end-to-end) below; the BAMs and CRAMs are most
  of it
- At least 32 GB of RAM for the minimap2 alignment step. Only data preparation
  needs it; a 2 GB instance hosts the finished site.
- The command-line tools below, with the versions tested in parentheses:
  - [JBrowse CLI](https://jbrowse.org/jb2-staging/docs/cli) (`@jbrowse/cli` v3.6.5 or later)
  - [Node.js](https://nodejs.org/) (v18 minimum, v24.1.0 used for this tutorial)
  - [tabix](http://www.htslib.org/doc/tabix.html) (v1.21 or later)
  - [samtools](http://www.htslib.org/) (v1.21 or later)
  - [minimap2](https://github.com/lh3/minimap2)
  - [megadepth](https://github.com/ChristopherWilks/megadepth) (v1.2.0 or
    later), for the coverage tracks
  - [HiFiCNV](https://github.com/PacificBiosciences/HiFiCNV) (v1.0 or later),
    for the binned depth track

## Where the data comes from

HG008, the C-GIAB matched tumor/normal pair, under NCBI BioProject PRJNA200694
on the C-GIAB FTP, one dated directory per group that ran an analysis on it,
with the assemblies in NIST's S3 bucket and the per-clone CNV calls rehosted
here.

The reference and the reads:

- the C-GIAB reference build (GRCh38 with decoys and masked regions):
  https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/release/references/GRCh38/GRCh38_GIABv3_no_alt_analysis_set_maskedGRC_decoys_MAP2K3_KMT2C_KCNJ18.fasta.gz
- the tumor/normal PacBio HiFi reads (Revio run, 116x tumor, 35x normal):
  https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/Liss_lab/PacBio_Revio_20240125/

The somatic call sets, one per group that published one:

- the V0.5 draft benchmark SV and CNV calls:
  https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/Liss_lab/analysis/NIST_HG008-T_somatic-stvar-CNV_DraftBenchmark_V0.5-20260318/
- Severus somatic SVs (HiFi):
  https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/Liss_lab/analysis/NIH_HiFi_Severus-SV_20240308/somatic_SVs/severus_somatic.vcf.gz
- the minda ensemble SVs (HiFi, ONT and Illumina callers):
  https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/Liss_lab/analysis/NIH-NCI_minda-ensemble_20240710/HG008_minda_ensemble.vcf
- DRAGEN's somatic SV and CNV calls (Illumina):
  https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/Liss_lab/analysis/DRAGEN-v4.2.4_ILMN-WGS_20240312/standard/
- NYGC's somatic SVs, annotated CNV segments and BIC-seq2 log2 ratio (Illumina):
  https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/Liss_lab/analysis/NYGC-somatic-pipeline_20240412/GRCh38-GIABv3/
- Wakhan's haplotype-specific copy number and LOH segments (HiFi phased with
  Hi-C):
  https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/Liss_lab/analysis/NIH_HiFi-HiC_Wakhan-CNA_20240424/bed_output/
- the earlier Wakhan run's Clair3 tumor small-variant calls, which HiFiCNV reads
  for its own MAF track:
  https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/Liss_lab/analysis/NIH_HiFi_Wakhan-CNA_20240308/vcf_inputs/merge_output_tumor.vcf.gz
- the normal's germline calls, which the BAF track piles the tumor reads up
  against:
  https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/Liss_lab/analysis/PacBio_Revio_20240125/pacbio-wgs-wdl_germline_20240206/HG008-N-P.GRCh38.deepvariant.phased.vcf.gz

The assemblies:

- the T2T tumor assembly, v3.2:
  https://nist-giab.s3.us-east-1.amazonaws.com/giab_tumor-normal/analysis/HG008/NIST_asm_dev/HG008T_v3.2/HG008T_v3.2.fasta.gz
- the matched normal assembly, v6.3:
  https://nist-giab.s3.us-east-1.amazonaws.com/giab_tumor-normal/analysis/HG008/NIST_asm_dev/HG008N_v6.3/HG008N_v6.3.fasta.gz

The single-cell-derived clone panel:

- short-read WGS, one run per clone:
  https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/NIST/HG008-T_clones/
- the per-clone CNVkit calls, rehosted merged into one multi-row BED:
  https://jbrowse.org/demos/cgiab/HG008T-clones.cnv.multirow.bed.gz

## The C-GIAB dataset

[Cancer Genome in a Bottle (C-GIAB)](https://www.nist.gov/programs-projects/cancer-genome-bottle)
publishes HG008 as a matched pair from one pancreatic ductal adenocarcinoma
(PDAC) donor: the tumor (HG008-T) and normal pancreatic tissue (HG008-N-P).

HG008-T is **hypodiploid**: its assembly recapitulates 35 tumor chromosomes,
down from 46, with widespread arm-level loss and 16 truncal interchromosomal
rearrangements
([Wagner et al. 2026](https://doi.org/10.64898/2026.05.01.722316)). Arm-level
loss is the backdrop for every copy-number figure below, so the benchmark CNV
BED, which reports absolute copy number per interval, is what every other track
gets read against.

The [SV visualization guide](https://jbrowse.org/jb2-staging/docs/user_guides/sv_visualization) and the
[SV inspector guide](https://jbrowse.org/jb2-staging/docs/user_guides/sv_inspector_view) cover the concepts;
this page is the data-loading workflow and a few worked examples. For the rest
of the call sets, the auxiliary assays and the methods, see the
[NIST C-GIAB page](https://www.nist.gov/programs-projects/cancer-genome-bottle)
and [McDaniel et al. 2025](https://doi.org/10.1038/s41597-025-05438-2).

## Setting up

The instance itself is the [web quickstart](https://jbrowse.org/jb2-staging/docs/quickstart_web) unchanged. Two
of the prerequisites install from release binaries:

```bash
wget https://github.com/ChristopherWilks/megadepth/releases/download/1.2.0/megadepth
chmod +x megadepth && sudo mv megadepth /usr/local/bin/
curl -L https://github.com/PacificBiosciences/HiFiCNV/releases/download/v1.0.1/hificnv-v1.0.1-x86_64-unknown-linux-gnu.tar.gz \
  | tar xz --strip-components=1 -C /usr/local/bin --wildcards '*/hificnv'
```

Load the C-GIAB build of GRCh38 as the assembly: it carries decoys and several
masked regions, and it is also the reference the CRAM conversion below writes
against.

[Reproduce it end to end](#reproduce-it-end-to-end) fetches that reference and
every file below in one script; the sections below give the track config for
each file.

## The benchmark SV and CNV calls

The V0.5 HG008-T draft benchmark is two files: the SV calls as an indexed VCF,
the CNV calls as a BED. Both load straight from their FTP URL with nothing
downloaded. The walkthroughs below open individual calls out of the SV track, so
it goes in first.

```json addtrack
{
  "type": "VariantTrack",
  "trackId": "hg008t_benchmark_sv",
  "name": "HG008-T V0.5 draft benchmark somatic SVs",
  "assemblyNames": ["GRCh38_GIABv3"],
  "adapter": {
    "type": "VcfTabixAdapter",
    "uri": "https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/Liss_lab/analysis/NIST_HG008-T_somatic-stvar-CNV_DraftBenchmark_V0.5-20260318/GRCh38_HG008-T-V0.5_somatic-stvar_PASS.draftbenchmark.vcf.gz"
  }
}
```

The CNV BED ships without a header, so its columns beyond `chrom/start/end` load
unnamed. Name them on the adapter with the
[`columnNames`](https://jbrowse.org/jb2-staging/docs/config/bedadapter/#slot-columnnames) slot:

```json addtrack
{
  "type": "FeatureTrack",
  "trackId": "hg008t_somatic_cnv",
  "name": "HG008-T somatic CNV",
  "assemblyNames": ["GRCh38_GIABv3"],
  "adapter": {
    "type": "BedAdapter",
    "uri": "https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/Liss_lab/analysis/NIST_HG008-T_somatic-stvar-CNV_DraftBenchmark_V0.5-20260318/GRCh38_HG008-T-V0.5_somatic-CNV_PASS.draftbenchmark.calls.bed",
    "columnNames": [
      "chrom",
      "start",
      "end",
      "total_copy_number",
      "hap1_copy_number",
      "hap2_copy_number",
      "name"
    ]
  }
}
```

## The reads and their coverage

The tumor and normal BAMs at the C-GIAB FTP are large, slow to read remotely,
and carry no `MD` tags, which JBrowse uses to display mismatches without
re-fetching the reference. Pull each one through `samtools view` into a local
CRAM against the reference above, and write a whole-genome coverage bigWig
beside it:

<!-- from: scripts/build_sv_visualization_cgiab.sh -->

```bash
# -T names the reference the CRAM is written against, and it has to be the same
# build the assembly was loaded from or every base reads as a mismatch
# --write-index saves a second samtools pass for the .crai
samtools view HG008-T.bam --write-index -o HG008-T.cram -T GRCh38.fa

# one whole-genome coverage bigWig per sample, written beside its CRAM
megadepth HG008-T.cram --bigwig
```

## Structural variants from the published callsets

The benchmark is one of five somatic SV callsets on this pair, and C-GIAB
publishes the rest as one URL each, loaded the way the benchmark was:

| Callset                                                                                                                 | Called from                                    |
| ----------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------- |
| NIST V0.5 draft benchmark                                                                                               | assembly comparison plus read support          |
| [Severus](https://github.com/KolmogorovLab/Severus)                                                                     | PacBio HiFi                                    |
| [minda](https://github.com/KolmogorovLab/minda) ensemble                                                                | eleven caller runs over HiFi, ONT and Illumina |
| DRAGEN                                                                                                                  | Illumina WGS                                   |
| NYGC somatic pipeline ([Manta](https://github.com/Illumina/manta) and [GRIDSS](https://github.com/PapenfussLab/gridss)) | Illumina WGS                                   |

<Figure caption="The chr3 breakends of the benchmark's cluster_3 in five SV callsets, over the HiFiCNV depth and the benchmark's CNV lane: the V0.5 benchmark, Severus, the minda ensemble, DRAGEN and NYGC's BEDPE. Every callset marks both breakends, the depth steps down between them, and the CNV lane crosses the whole window as one segment." src="/img/sv_cgiab/sv_callset_comparison.png" />

The benchmark files the two chr3 breakends in that window under one `EVENT`,
`cluster_3`, while its CNV BED covers the whole window with a single segment
named `noCNV`. The CNV callset works at a coarser scale, so an event this size
is the SV lanes' to carry.

The second of those breakends is written two ways. The benchmark and minda place
a breakend there; Severus and DRAGEN write a symbolic inversion whose `SVLEN`
runs far down the arm, so those two lanes draw a span leaving the window where
the others draw a mark.

### Severus and DRAGEN

Both are indexed VCFs with a record at each breakend, so they load with no
display settings:

```json addtrack
{
  "type": "VariantTrack",
  "trackId": "hg008t_severus_sv",
  "name": "HG008-T Severus somatic SVs (HiFi)",
  "assemblyNames": ["GRCh38_GIABv3"],
  "adapter": {
    "type": "VcfTabixAdapter",
    "uri": "https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/Liss_lab/analysis/NIH_HiFi_Severus-SV_20240308/somatic_SVs/severus_somatic.vcf.gz"
  }
}
```

```json addtrack
{
  "type": "VariantTrack",
  "trackId": "hg008t_dragen_sv",
  "name": "HG008-T DRAGEN somatic SVs (Illumina)",
  "assemblyNames": ["GRCh38_GIABv3"],
  "adapter": {
    "type": "VcfTabixAdapter",
    "uri": "https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/Liss_lab/analysis/DRAGEN-v4.2.4_ILMN-WGS_20240312/standard/dragen_4.2.4_HG008-mosaic_tumor.sv.vcf.gz"
  }
}
```

DRAGEN's record over the chr3-chr13 junction places the breakend where the
others do and carries `MaxDepth` in its own `FILTER` column, the depth cap a
caller applies where a pileup runs far above the genome average. Clicking the
record shows it.

### minda: the caller runs behind each junction

The ensemble callset is a plain VCF, so [`VcfAdapter`](https://jbrowse.org/jb2-staging/docs/config/vcfadapter)
loads it whole, with no index:

```json addtrack
{
  "type": "VariantTrack",
  "trackId": "hg008t_minda_sv",
  "name": "HG008-T minda ensemble SVs (HiFi, ONT, Illumina)",
  "assemblyNames": ["GRCh38_GIABv3"],
  "adapter": {
    "type": "VcfAdapter",
    "uri": "https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/Liss_lab/analysis/NIH-NCI_minda-ensemble_20240710/HG008_minda_ensemble.vcf"
  }
}
```

Clicking a junction shows its `SUPP_VEC`, which names every caller run that
supported it, each prefixed by the technology it ran on: `PB_severus_BND1941_1`,
`ONT_Sniffles2.INV.30AM2`, `ILL_gridss63ff_2860h`.

### NYGC: a BEDPE drawn as arcs

[`BedpeAdapter`](https://jbrowse.org/jb2-staging/docs/config/bedpeadapter) reads a paired-end BED whole, with
no index, and serves it to a variant track. The `#` header names the columns, so
each record arrives with the pipeline's call, the strands, the CNV changepoints
it was linked to, and an `evidence` column holding each tool's split-read and
paired-end counts:

```json addtrack
{
  "type": "VariantTrack",
  "trackId": "hg008t_nygc_sv",
  "name": "HG008-T NYGC somatic SVs (Manta, GRIDSS)",
  "assemblyNames": ["GRCh38_GIABv3"],
  "adapter": {
    "type": "BedpeAdapter",
    "uri": "https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/Liss_lab/analysis/NYGC-somatic-pipeline_20240412/GRCh38-GIABv3/HG008-T--HG008-N.sv.annotated.v7.somatic.high_confidence.final.bedpe"
  }
}
```

Each record holds both ends, so the whole callset reads as arcs at chromosome
scale: **Display types → Variant display arcs** on the track menu draws one arc
per record between its two breakends.

## Copy number from the published callsets

Four groups have called copy number on this pair, and C-GIAB publishes each
one's output. Every file is small and loads from its FTP URL, so they can share
one view:

| Callset                                                                                                                                  | Called from                           | Each segment carries                                                        |
| ---------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------- | --------------------------------------------------------------------------- |
| NIST V0.5 draft benchmark                                                                                                                | assembly comparison plus read support | absolute total and per-haplotype copy number                                |
| [Wakhan](https://github.com/KolmogorovLab/Wakhan)                                                                                        | PacBio HiFi, phased with Hi-C         | copy number per parental haplotype, with LOH intervals in a second file     |
| NYGC somatic pipeline, [BIC-seq2](https://doi.org/10.1073/pnas.1110574108)                                                               | Illumina WGS                          | log2 tumor-versus-normal copy ratio, and the genes the segment covers       |
| [DRAGEN](https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/Liss_lab/analysis/DRAGEN-v4.2.4_ILMN-WGS_20240312/) | Illumina WGS                          | integer copy number, minor-haplotype copy number and minor allele frequency |

The benchmark CNV BED added above is the lane the others get read against, since
its copy numbers are absolute: it states a diploid region explicitly as CN 2,
and it leaves out segments whose breakpoints the project could not place on
GRCh38, so a gap in that lane is a gap in the benchmark. Depth per bin is the
one signal no group publishes; the end of this section computes it from the
tumor reads.

<Figure caption="Four published CNV callsets over chr9p21.3, with the HiFiCNV depth and B-allele frequency above them: the V0.5 benchmark, NYGC's annotated BIC-seq2 segments, DRAGEN's somatic CNV VCF, and Wakhan's two haplotype rows. Depth drops out over CDKN2A, where the benchmark and NYGC both carry a focal call and the two coarser segmentations run straight through." src="/img/sv_cgiab/cnv_callset_comparison.png" />

### DRAGEN: integer copy number from short reads

Each record in DRAGEN's somatic CNV VCF is one segment, and its ID names the
class the caller assigned: `LOSS`, `GAIN`, `CNLOH` for copy-neutral loss of
heterozygosity, and `REF` for a segment left at the reference state.

```json addtrack
{
  "type": "VariantTrack",
  "trackId": "hg008t_dragen_cnv",
  "name": "HG008-T DRAGEN somatic CNV (Illumina)",
  "assemblyNames": ["GRCh38_GIABv3"],
  "adapter": {
    "type": "VcfTabixAdapter",
    "uri": "https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/Liss_lab/analysis/DRAGEN-v4.2.4_ILMN-WGS_20240312/standard/dragen_4.2.4_HG008-mosaic_tumor.cnv.vcf.gz"
  }
}
```

The VCF header carries DRAGEN's purity and ploidy estimate for the sample, and
its ploidy lands below two: the hypodiploidy the assembly reports, arrived at
from short reads by another caller.

### NYGC: a copy ratio, and the genes each segment covers

The New York Genome Center's somatic pipeline ran on the same pair, and C-GIAB
publishes its CNV output two ways. `HG008-T--HG008-N.cnv.annotated.v7.final.bed`
needs nothing done to it: its `#` header line is tab-separated, so the adapter
takes the column names from the file.

```json addtrack
{
  "type": "FeatureTrack",
  "trackId": "hg008t_nygc_cnv",
  "name": "HG008-T NYGC CNV calls, annotated (BIC-seq2)",
  "assemblyNames": ["GRCh38_GIABv3"],
  "adapter": {
    "type": "BedAdapter",
    "uri": "https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/Liss_lab/analysis/NYGC-somatic-pipeline_20240412/GRCh38-GIABv3/HG008-T--HG008-N.cnv.annotated.v7.final.bed"
  },
  "displayDefaults": {
    "color": "jexl:feature.type=='DEL'?'#2166ac':'#b2182b'",
    "labels": { "name": "jexl:feature.type+' '+feature.cytoband" },
    "displayMode": "compact",
    "legend": [
      { "label": "Loss (DEL)", "color": "#2166ac" },
      { "label": "Gain (DUP)", "color": "#b2182b" }
    ]
  }
}
```

Clicking a segment then shows its call, its log2 ratio, its cytoband, the focal
flag, and each Cancer Gene Census gene at its tier.

`HG008-T--HG008-N.bicseq2.txt` is the same segmentation in quantitative form,
one log2 ratio per segment, and one `awk` away from a bedGraph:

<!-- from: scripts/build_sv_visualization_cgiab.sh -->

```bash
# column 9 is log2.copyRatio, and the file is 1-based where bedGraph is not
awk 'NR>1 {printf "%s\t%d\t%d\t%.4f\n", $1, $2-1, $3, $9}' \
  HG008-T--HG008-N.bicseq2.txt > HG008-T_bicseq2_log2ratio.bedgraph
```

Plot it as a **Line (step)** over a fixed range, since a homozygous deletion
carries no reads and so no finite ratio. The balanced baseline sits above zero
because BIC-seq2 normalizes on total read counts and this genome is hypodiploid;
read its steps against the benchmark's absolute copy numbers.

### Wakhan: copy number per parental haplotype

A log2 ratio and a folded allele frequency both average the two parental alleles
together, so at whole-genome zoom an LOH block reads as balanced.
[Wakhan](https://github.com/KolmogorovLab/Wakhan) phases the germline
heterozygous SNPs and reports copy number _per haplotype_, which keeps the LOH
signal at every zoom. C-GIAB publishes two Wakhan runs on this tumor, and the
later one phases the normal with Arima Hi-C alongside the HiFi reads:
`HG008_HiFi_HiC_copynumbers_segments.bed` and `HG008_HiFi_HiC_loh_segments.bed`,
both of them URL tracks.

The copy-number file is worth a few more lines of config than `add-track`
writes. It is long format, one row per haplotype, and its column-name line
carries no `#`:

```
chr	start	end	copynumber_state	coverage	haplotype
chr1	0	23750000	2	106.025	1
chr1	23750001	119650000	0.72	58.025	1
```

so name the columns on the adapter with
[`columnNames`](https://jbrowse.org/jb2-staging/docs/config/bedadapter/#slot-columnnames). The header line
itself loads as a feature on a refName no assembly has, so nothing draws it.

Because a haplotype column already assigns each segment to a row, this is a
[`LinearMultiRowFeatureDisplay`](https://jbrowse.org/jb2-staging/docs/config/linearmultirowfeaturedisplay)
track: set
[`partitionField`](https://jbrowse.org/jb2-staging/docs/config/linearmultirowfeaturedisplay/#slot-partitionfield)
to `haplotype` and it paints one row per parental copy.

```json
{
  "type": "FeatureTrack",
  "trackId": "hg008_wakhan_haplotype",
  "name": "HG008-T Wakhan copy number per haplotype",
  "assemblyNames": ["GRCh38_GIABv3"],
  "adapter": {
    "type": "BedAdapter",
    "uri": "https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/Liss_lab/analysis/NIH_HiFi-HiC_Wakhan-CNA_20240424/bed_output/HG008_HiFi_HiC_copynumbers_segments.bed",
    "columnNames": [
      "chrom",
      "start",
      "end",
      "copynumber_state",
      "coverage",
      "haplotype"
    ]
  },
  "displays": [
    {
      "type": "LinearMultiRowFeatureDisplay",
      "displayId": "hg008_wakhan_haplotype-LinearMultiRowFeatureDisplay",
      "partitionField": "haplotype",
      "color": "jexl:get(feature,'copynumber_state')<0.5?'#2166ac':get(feature,'copynumber_state')<1.5?'#bdbdbd':'#f4a582'",
      "legend": [
        { "label": "Haplotype lost (0)", "color": "#2166ac" },
        { "label": "One copy", "color": "#bdbdbd" },
        { "label": "Two or more copies", "color": "#f4a582" }
      ]
    }
  ]
}
```

`copynumber_state` counts one parental copy, so `1` is the expected state and a
`0` row is the lost haplotype that makes an arm LOH. Wakhan emits fractional
states for segments that are not clonal, which is why the color buckets ranges.
`coverage` is Wakhan's median depth for the segment, so the per-copy depth scale
can be read straight off it.

### Depth per bin, and B-allele frequency

The coverage bigWigs above are raw depth. Two more tracks, both built from the
tumor reads, are what make copy number readable beside the callsets: binned
depth, and B-allele frequency.
[HiFiCNV](https://github.com/PacificBiosciences/HiFiCNV), PacBio's somatic CNV
caller, writes the first:

<!-- from: scripts/build_sv_visualization_cgiab.sh -->

```bash
# --maf holds the TUMOR's small-variant calls, the Clair3 tumor VCF published
# alongside C-GIAB's Wakhan run: HiFiCNV reads AD out of this VCF for its own
# allele-frequency output and never looks at --bam for it
hificnv --bam HG008-T.cram --ref GRCh38.fa --maf tumor_smallvariants.vcf.gz \
  --output-prefix hificnv
```

HiFiCNV names each output for the sample it came from, so the depth bigWig
carries the `--bam` sample's name. Give it the **Scatter** plot type: depth is a
read count per bin, so a whole-chromosome view is a cloud hundreds of points
deep and a copy-number step is wherever its centre moves. The NYGC copy ratio
above is that same signal already segmented.

The allelic panel is **B-allele frequency**, unfolded. HiFiCNV's own `maf.bw`
folds to `min(AF, 1-AF)`, so a region that has lost one parental copy collapses
onto a single band near 0. Unfolded BAF keeps the two apart: a balanced region
is one band at 0.5, a loss-of-heterozygosity region two bands at 0 and 1. Build
it by piling up the tumor reads at the sites the **normal** calls heterozygous
and taking the alt fraction:

<!-- from: scripts/build_sv_visualization_cgiab.sh -->

```bash
# het sites from the NORMAL, which is the choice the whole track rests on: an LOH
# site is homozygous in the tumor, so a tumor-derived list drops exactly the
# sites the track exists to show
bcftools view -g het -Oz -o hets.vcf.gz normal.deepvariant.vcf.gz
tabix -p vcf hets.vcf.gz
cut -f1,2 GRCh38.fa.fai > GRCh38.chrom.sizes

# -q 1 drops multi-mapped reads, -Q 0 leaves HiFi base qualities alone
bcftools mpileup -f GRCh38.fa -T hets.vcf.gz -a AD -q 1 -Q 0 tumor.bam |
  bcftools query -f '%CHROM\t%POS\t[%AD]\n' |
  # unfolded alt fraction, so LOH separates into 0 and 1 instead of folding
  # onto one band; the 10x floor keeps thin coverage from painting a fake 0/1
  awk -F'[\t,]' '{d=$3+$4; if (d>=10) printf "%s\t%d\t%d\t%.4f\n",$1,$2-1,$2,$4/d}' |
  LC_COLLATE=C sort -k1,1 -k2,2n > baf.bedgraph
bedGraphToBigWig baf.bedgraph GRCh38.chrom.sizes tumor_baf.bw
```

Plot it with **Scatter** over a fixed 0 to 1 range: the spread is the entire
signal.

<Figure caption="Chromosome 3 over the benchmark CNV calls: BIC-seq2's segmented log2 copy ratio, the HiFiCNV depth it summarizes, and B-allele frequency. The p-arm is a single-copy loss with loss-of-heterozygosity; the q-arm is balanced." src="/img/sv_cgiab/cnv_depth_baf.png" />

#### Keep the BAF track off bigWig summaries

A bigWig carries precomputed zoom levels, and each zoomed bin holds a minimum,
an average and a maximum. BAF is a _distribution_, so every bin over an LOH arm
comes back as minimum 0, maximum 1 and an average that wanders. The default
[`summaryScoreMode`](https://jbrowse.org/jb2-staging/docs/config/linearwiggledisplay/#slot-summaryscoremode) of
`whiskers` draws all three, which paints the arm as a solid full-height wash.

The fix goes on the adapter, with
[`resolutionMultiplier`](https://jbrowse.org/jb2-staging/docs/config/bigwigadapter/#slot-resolutionmultiplier).
It scales the bases-per-bin the adapter asks for, and a small enough value keeps
the fetch on the raw per-site values at the zoom levels these figures use:

```json
{
  "type": "QuantitativeTrack",
  "trackId": "HG008-T_baf",
  "name": "HG008-T B-allele frequency (BAF)",
  "assemblyNames": ["GRCh38_GIABv3"],
  "adapter": {
    "type": "BigWigAdapter",
    "uri": "HG008-T_baf.bcftools.bw",
    "resolutionMultiplier": 0.001
  },
  "displayDefaults": {
    "defaultRendering": "scatter",
    "scatterPointSize": 1,
    "minScore": 0,
    "maxScore": 1
  }
}
```

Raw het sites are cheap here, because the track carries one value per germline
heterozygous site rather than one per base. Whole-genome view still falls back
to the summary, which is what Wakhan's segments above are for.

**Resolution → Finer** in the track menu is the same control interactively.
Reach for it whenever a scatter track paints as a filled band.

### Subclonal copy number

Every callset above averages over the cells it was sequenced from, so a change
only part of the tumor carries reads as a muted, intermediate signal. HG008-T is
such a mixture: karyotyping across passages finds the arm-level losses in nearly
every cell, and finds the genome-doubled fraction of the culture growing between
early and late passage
([Wagner et al. 2026](https://doi.org/10.64898/2026.05.01.722316)). The
benchmark CNV BED reports copy number for the cells that have not doubled.

C-GIAB publishes short-read WGS for a panel of HG008-T single-cell-derived
clones under
[`HG008-T_clones/`](https://ftp-trace.ncbi.nlm.nih.gov/ReferenceSamples/giab/data_somatic/HG008/NIST/HG008-T_clones/).
Each clone is a colony grown from one tumor cell, so each reports one subclone's
copy number. Called per clone and merged into one BED with a `clone` column,
they partition into rows the way the Wakhan haplotypes do, and a row that
departs from the rest is a CNV private to that subclone:

```json addtrack
{
  "type": "FeatureTrack",
  "trackId": "hg008_subclonal_cnv",
  "name": "HG008-T subclonal CNV (per-clone CNVkit)",
  "assemblyNames": ["GRCh38_GIABv3"],
  "adapter": {
    "type": "BedTabixAdapter",
    "uri": "https://jbrowse.org/demos/cgiab/HG008T-clones.cnv.multirow.bed.gz"
  },
  "displays": [
    {
      "type": "LinearMultiRowFeatureDisplay",
      "displayId": "hg008_subclonal_cnv-LinearMultiRowFeatureDisplay",
      "partitionField": "clone",
      "color": "jexl:get(feature,'cn')<1?'#2166ac':get(feature,'cn')<2?'#92c5de':get(feature,'cn')<3?'#e0e0e0':get(feature,'cn')<4?'#f4a582':'#b2182b'",
      "legend": [
        { "label": "CN 0", "color": "#2166ac" },
        { "label": "CN 1", "color": "#92c5de" },
        { "label": "CN 2", "color": "#e0e0e0" },
        { "label": "CN 3", "color": "#f4a582" },
        { "label": "CN 4+", "color": "#b2182b" }
      ]
    }
  ]
}
```

Read those integers on the caller's own scale: CNVkit centers each sample's log2
on that sample's own median, so on a hypodiploid genome the balanced state is
not the row's CN 2. The benchmark CNV track in the same view is what anchors
them, its `total_copy_number` being absolute.

The p-arm of chr3 reads clearly for that reason: the benchmark calls one state
across the whole arm, and the bulk depth holds one level under it.

<Figure caption="The p-arm of chr3 over the HG008-T clones: the HiFiCNV depth, the benchmark CNV call, and one row per clone from the per-clone CNVkit BED. One row departs from the rest at the p-terminus and rejoins them partway down the arm." src="/img/sv_cgiab/subclonal_cnv.png" />

## Align the tumor assembly to GRCh38

The tumor assembly is haplotype-resolved into T2T scaffolds. Load it as a second
JBrowse assembly and align it to GRCh38; the PAF that comes out is what the
synteny and dotplot views draw:

<!-- from: scripts/build_sv_visualization_cgiab.sh -->

```bash
# asm5 is the same-species preset; -c emits the base-level CIGAR the synteny
# view needs to draw a junction at base scale
minimap2 -cx asm5 GRCh38.fa HG008T_v3.2.fasta > HG008T_v3.2.paf

# -a is query,target, the REVERSE of the target query minimap2 just took: get it
# backwards and the view opens empty rather than erroring
jbrowse add-track HG008T_v3.2.paf -a HG008T_v3.2,GRCh38_GIABv3
```

The matched normal assembly (`HG008N_v6.3.fasta.gz`, same S3 path) loads the
same way. See the
[synteny track config guide](https://jbrowse.org/jb2-staging/docs/config_guides/synteny_track) and the
[linear synteny view guide](https://jbrowse.org/jb2-staging/docs/user_guides/linear_synteny_view).

## Walkthroughs

The loaded data reads three complementary ways: the SV inspector for
whole-genome triage, the linear genome view for read-level detail and copy
number, and the dotplot and synteny views for chromosome-scale rearrangements in
the assembly. Each walkthrough below runs on an instance built from the sections
above, or on
[the hosted C-GIAB demo](https://jbrowse.org/code/jb2/main/?config=https://jbrowse.org/demos/cgiab/config.json),
which carries the benchmark calls, the reads and the copy-number tracks already
loaded.

### A chr3-chr13 translocation

**Add → SV inspector**, then **Open from track** to pick the C-GIAB benchmark
VCF loaded earlier.

<Figure caption="The SV inspector showing the benchmark VCF as a circular overview alongside a table of calls." src="/img/sv_cgiab/translocation_sv_inspector_view.png" />

Click the chord joining chr3 and chr13, then open the tumor PacBio HiFi reads on
each panel of the breakpoint split view it launches and set **Read height** →
**Compact**.

<Figure caption="Clicking the chord joining chr3 and chr13 opens a breakpoint split view. Black splines connect tumor PacBio HiFi reads that partially map to each chromosome, suggesting a fusion or translocation." src="/img/sv_cgiab/translocation_breakpoint_split.png" />

That junction is one breakend of a larger event, and interchromosomal
translocations in HG008 are frequently complex this way
([Wagner et al. 2026](https://doi.org/10.64898/2026.05.01.722316)). The next
walkthrough reads this one three ways.

For the SV inspector workflow itself (filtering the table, search, configuring
the circular overview), see the
[SV inspector guide](https://jbrowse.org/jb2-staging/docs/user_guides/sv_inspector_view).

### The same junction three ways

The chord says where to look. Three things in this instance say what is there,
and none of them is derived from the others.

**The caller.** `SV_20` and `SV_190` are one junction written twice, joining
chr3:139,976,414 to chr13:114,353,244. A BND record names one partner, so each
describes a translocation. The `EVENT` field groups them: the benchmark files
both under `cluster_3` alongside two further breakends and tags them
`EVENTTYPE=CHROMOPLEXY`. A caller can group junctions into an event because it
sees the whole callset at once. It cannot say which molecule carries them.

**The reads.** Put both breakpoint loci on screen and, from the tumor PacBio
HiFi track's menu, choose **Launch → Reconstruct derivative allele...**. The
reads in the window are grouped by the route their split alignments describe,
each offered with the number of reads that independently take it. The top route
runs chr13 forward into the junction and then down chr3 inverted, the
orientation the black splines above draw. The matched normal is the control and
a track away: the tumor reads split at this position, the normal reads read
through it.

<Figure caption="Reconstruct derivative allele over both breakpoint loci of the tumor PacBio HiFi track. The top route, chr13 forward then chr3 inverted, is the junction the benchmark and the tumor assembly both name." src="/img/sv_cgiab/three_ways.png" />

<Video src="/media/sv/derivative_allele_route.mp4" caption="The same route end to end, on the COLO829 melanoma chain: the track menu, the ranked routes, and Draw as Breakpoint split view replacing the window with one panel per segment of the route the reads take." />

**The assembly.** The synteny track loaded earlier says the same thing from no
reads at all. The C-GIAB assembly resolves both loci onto a single tumor contig,
and named that contig for the two chromosomes it fuses. Its chr13 arm ends at
the chr13 breakend above and its chr3 arm begins at the chr3 one, abutting at a
single base of contig coordinate, with the same orientation flip the reads
describe. Open it in the synteny or dotplot view against GRCh38 and the junction
is the point where one contig stops following chr13 and starts following chr3.

Back in the reconstruction, reading the list below the top route is the other
half of the exercise. This window ends at the chr13 q-terminus, so most of what
is offered under the real junction is reads mismapped into the terminal repeats
of other chromosomes, each a confident-looking two-segment route with a real
read count behind it. The read count ranks the routes. What picks this one out
is that the caller and the assembly put its two ends in the same two places.

The reconstruction is bounded twice by what is loaded. It is assembled from the
reads in the **displayed regions**, which is why both sides of this junction are
open above. And the hosted demo slices the tumor reads to the loci these
walkthroughs visit, so the reads reach one of `cluster_3`'s junctions where the
assembly contig carries both. Rebuilding from the full BAM with
[the build script](#reproduce-it-end-to-end) lifts that limit.

### Which calls are drivers

Most somatic calls in a tumor genome are passengers: real events, carried along
by the cell lineage, with no role in the cancer. A handful are drivers. In
pancreatic ductal adenocarcinoma the recurrently altered genes are _KRAS_,
_CDKN2A_, _TP53_ and _SMAD4_
([Waddell et al. 2015](https://doi.org/10.1038/nature14169),
[Bailey et al. 2016](https://doi.org/10.1038/nature16965)), and the copy-number
walkthroughs below visit all four in this genome. The deletion that comes first
is a passenger, and reads exactly the same way.

The benchmark BED states copy number and haplotype; consequence comes from a
driver catalogue, and the somatic ones (COSMIC's Cancer Gene Census among them)
are licensed, so a public demo cannot carry one as a lane. The NYGC segments
loaded above carry theirs inline, each one listing the census genes inside it.
Every copy-number figure below also draws one MANE Select transcript under the
lanes, so the event and the gene it covers are read off the same axis.

### A small deletion in CUZD1

For small to medium SVs the linear genome view is usually enough. Use the
**search** (magnifying glass) button in the SV inspector to find a specific
call, for example `SV_85`, a heterozygous deletion that affects two exons of
_CUZD1_.

_CUZD1_ is a passenger here, a pancreatic acinar protein predicted to act in
trypsinogen activation
([NCBI Gene 50624](https://www.ncbi.nlm.nih.gov/gene/50624)). The call takes one
of its two copies, and at ~1.8 kb over two exons it reads base by base in a
pileup.

Whether anyone has submitted a CNV here is a lane of its own: **ClinVar CNVs**
carries the submitted copy-number variants and their clinical significance. Add
it from UCSC:

```json addtrack
{
  "type": "FeatureTrack",
  "trackId": "hg38_clinvar_cnv_ucsc",
  "name": "ClinVar CNVs (UCSC)",
  "assemblyNames": ["GRCh38_GIABv3"],
  "adapter": {
    "type": "BigBedAdapter",
    "uri": "https://hgdownload.soe.ucsc.edu/gbdb/hg38/bbi/clinvar/clinvarCnv.bb"
  },
  "displayDefaults": {
    "jexlFilters": ["get(feature,'_varLen') < 50000"],
    "displayMode": "compact"
  }
}
```

ClinVar holds chromosome-scale records alongside focal ones, so the size filter
keeps the lane at this event's scale; `_varLen` is the catalogue's own length
field. The lane comes back empty over this locus, no submitted CNV near this
deletion's size covering it.

<Figure caption="The SV inspector after searching for SV_85, a heterozygous CUZD1 deletion, and the linear genome view its location link opens: the <DEL> ALT allele over the ClinVar CNV and NCBI RefSeq gene lanes." src="/img/sv_cgiab/deletion_sv_inspector_search.png" />

Open the gene annotations and the tumor PacBio HiFi reads, set **Read height →
Compact** and **Sort by... → Base pair** from the track menu, and center the
deletion. Turning on the view menu's **center line** helps line the breakpoint
up.

<Figure caption="Tumor PacBio HiFi reads at compact height, sorted by base pair with the deletion centered, over the gene annotations. The deletion removes two CUZD1 exons and is heterozygous." src="/img/sv_cgiab/deletion_linear_view.png" />

### A tandem-repeat call, sized against the normal

Some benchmark records are sized against the donor's own germline allele:
`EVENTTYPE=CNV:TR`, a somatic size change inside a tandem repeat, taken from the
tumor assembly aligned to the normal assembly. Their `SVLEN` is a number GRCh38
does not carry.

`SV_223` on chr5 is the benchmark's own worked example of one. Open it with both
samples' PacBio HiFi reads and sort each pileup at the call:

<Figure caption="SV_223 at base level: the benchmark's deletion call over the tumor and matched normal PacBio HiFi pileups, both sorted at the deleted span. The tumor's reads carry a deletion where the normal's carry an insertion at the same repeat, and the called span is wider than the deletion under it." src="/img/sv_cgiab/vntr_tumor_normal.png" />

`SVLEN` here is the span between the two samples' alleles rather than either
one's distance from the reference. The record states itself in the normal
assembly's coordinates as well, in `CHROM_HG8N6.3`, `POS_HG8N6.3`, `REF_HG8N6.3`
and `ALT_HG8N6.3`, so clicking it shows the same variant on the yardstick it was
called against.

Two more fields under it are worth the click. `SVVIZ_VAF_ALL` is the variant
allele fraction svviz2 measured across the HG008-T datasets, and
`SVVIZBYDATASET` breaks it into the sequencing runs behind it, which are worth
reading against each other at a repeat. The field's own description in the VCF
header says the estimate is biased there, and the pileup shows one reason why:
reads crossing a repeat spread the same allele over several alignments of it.

### Reading copy number

The quickest copy-number check is the tumor and normal coverage bigWigs as one
multi-bigwig track, which is fast at any zoom:

- **Show all regions in assembly** on the linear genome view start screen opens
  every chromosome at once.
- **Score → Set min/max score...** in the track menu pins the axis, holding the
  scale off the centromere and repeat spikes.
- **Plot type → Overlapping → Scatter** draws the two samples as points in one
  band, tumor red and normal blue.

<Figure caption="The linear genome view start screen: click Show all regions in assembly to lay out every chromosome across the view." src="/img/sv_cgiab/cnv_show_all_regions.png" />

The two rows in the figures here come from
[goleft indexcov](https://github.com/brentp/goleft/tree/master/indexcov), which
divides each sample by its own median. That normalization is what lets the rows
share an axis at all: the tumor and the normal were sequenced to different
depths, so raw coverage separates them before any copy number does, where
normalized rows put the normal flat at 1 and read every level the tumor holds as
a ratio against it. The pipeline above does not build those two files. They are
published beside the demo, as `HG008-N_indexcov.bw` and `HG008-T_indexcov.bw`
under https://jbrowse.org/demos/cgiab/, and load as a multi-wiggle track by URL.

Zoom to a region and open the benchmark CNV BED to check the coverage changes
against the called intervals. Coverage says a level changed; the BAF track in
the same window says what changed.

<Video src="/media/sv_cgiab/copy_number_layout.mp4" caption="Both menu routes on the coverage track, over chr5: Set min/max score pinning the axis, then Plot type to Overlapping Scatter, which redraws the two stacked rows as one band of points with the normal flat under the tumor's steps." />

<Figure caption="Chromosome 5: the segmented copy ratio, tumor and normal indexcov coverage as overlapping scatter, B-allele frequency, and the benchmark CNV calls. The normal stays flat while the tumor steps, and the BAF lane says what each step is." src="/img/sv_cgiab/cnv_with_bed_track.png" />

Four loci in HG008-T sit in four different copy-number states, and the depth,
BAF and copy-number lanes built above tell them apart:

| Locus  | State in HG008-T                 | Signature on the tracks             |
| ------ | -------------------------------- | ----------------------------------- |
| CDKN2A | Focal homozygous deletion (CN 0) | depth to 0, copy number 0           |
| TP53   | 17p loss + LOH (CN 1, 1+0)       | depth halved, BAF splits to 0 and 1 |
| SMAD4  | 18q loss + LOH (CN 1, 0+1)       | depth halved, BAF splits to 0 and 1 |
| KRAS   | Tandem duplication (CN 3, 2+1)   | depth raised, BAF to 1/3 and 2/3    |

All four carry clonal somatic variants in HG008
([Wagner et al. 2026](https://doi.org/10.64898/2026.05.01.722316)).

Arm-level loss is widespread in this hypodiploid genome, so a single band at 0.5
is the exception. Where a whole chromosome is LOH end to end, as chr17 is below,
the balanced band has to come from another chromosome.

#### CDKN2A: a homozygous deletion inside a single-copy loss

Navigate to `CDKN2A` on chr9, where the benchmark calls a focal ~20 kb
homozygous deletion over the gene (`SV_75`, total copy number 0). Both parental
copies are gone, so depth goes to ~0 under it. The deletion sits inside a larger
single-copy-loss arm (`CNA_14`, 0+1) where depth is already halved, so the focal
event removes the one copy the arm-level loss had left
([Wagner et al. 2026](https://doi.org/10.64898/2026.05.01.722316)).

Load the tumor and matched normal per-base coverage as one
[multi-quantitative track](https://jbrowse.org/jb2-staging/docs/user_guides/multiquantitative_track), one row
per sample, and set an explicit score range from the track menu so both rows are
drawn on the same scale. HiFiCNV's depth is binned where these two are per base,
and the PacBio HiFi read pileup below them is where the exact breakpoints are:
the thin lines crossing the gap in that pileup are single reads carrying the
deletion as one gap in their alignment.

The benchmark's `total_copy_number` is absolute, so CN 2 is a diploid segment.
The whole of 9p has lost a copy in this tumor, so CN 1 is the local background
here and the deletion is punched into it. Widen the view several hundred
kilobases to the right to reach the first CN 2 segment and read the CN 1 lane
against it.

<Figure caption="The CDKN2A deletion at 60 kb: coverage drops out in the tumor row and not in the normal, the read pileup drops out with it, and the CNV call under them reads CN 0." src="/img/sv_cgiab/driver_cdkn2a_deletion.png" />

#### chr17: loss with LOH, and copy-neutral LOH

Chromosome 17 carries a different LOH state on each arm. Open the whole
chromosome with the depth track above the BAF:

- the p-arm (covering _TP53_) is a single-copy loss with LOH (`CNA_20`, CN 1,
  1+0): depth is halved and the BAF splits away from 0.5.
- the q-arm is copy-neutral LOH (`CNA_21`, CN 2, 2+0): one parental haplotype
  was lost and the other duplicated, so total copy number is still 2 and depth
  stays flat, yet the BAF still splits away from 0.5.

The copy-ratio lane in the figure fills from a pivot at zero, and on this
hypodiploid genome the balanced level sits above zero, so the q-arm's
copy-neutral state fills upward in the gain color. Read that lane for where its
steps fall, not for which way they point.

<Figure caption="Chromosome 17: the segmented copy ratio, the HiFiCNV depth, the BAF and the benchmark CNV calls. The p-arm is a single-copy loss with LOH; the q-arm is copy-neutral LOH, flat in both copy-number lanes and still split in the BAF." src="/img/sv_cgiab/cnv_chr17_loh.png" />

The depth and BAF combinations read as a compact decision table:

| depth       | BAF             | Interpretation            |
| ----------- | --------------- | ------------------------- |
| flat (CN 2) | one band at 0.5 | balanced diploid          |
| flat (CN 2) | split to 0, 1   | copy-neutral LOH          |
| halved      | split to 0, 1   | single-copy loss with LOH |
| raised      | 1/3 and 2/3     | allelic gain              |

The benchmark BED's per-haplotype columns (`hap1_copy_number`,
`hap2_copy_number`) encode this allelic state: any segment with a `0` haplotype
(e.g. `1+0`, `2+0`) has lost one parental allele and its BAF splits away from
0.5, regardless of its total copy number. Clicking a CNV feature shows both.

#### KRAS and SMAD4

The same reading covers the other two loci. _KRAS_ on chr12 sits in a gain
(`SV_101`, CN 3, 2+1): the assembly resolves it as a 2 Mb tandem duplication
carrying the G12V-mutated copy, an event associated with advanced disease
([Wagner et al. 2026](https://doi.org/10.64898/2026.05.01.722316)). The event is
a couple of megabases, so zoom to it: at whole-chromosome scale it is a handful
of pixels wide.

<Figure caption="KRAS on chr12: its MANE Select transcript over the segmented copy ratio, the HiFiCNV depth and the BAF, above the CNV calls. Over the tandem duplication the copy-ratio edges land on the called boundaries and the BAF separates into two bands." src="/img/sv_cgiab/driver_kras_gain.png" />

_SMAD4_ on 18q is lost with LOH (`CNA_48`, CN 1, 0+1), the mirror image of the
_TP53_ event. Two controls are in the same picture: the balanced p-arm, and the
matched normal on the same axis as the tumor.

Leave the display's bicolor mode on and give it a symmetric axis, so the lane
fills from its zero pivot with a step down drawn the same size as a step up.

<Figure caption="Chromosome 18: SMAD4's MANE Select transcript over the segmented copy ratio, the tumor and its matched normal from indexcov, and the BAF, above the CNV calls. All three lanes change together from ~30 Mb to the telomere." src="/img/sv_cgiab/driver_smad4_loh.png" />

### Synteny and dotplot views of the tumor assembly

Side by side with the reference, the tumor assembly draws a complex SV as a
break in a diagonal. **Add → Dotplot view**, set the de novo assembly as one
axis and GRCh38 as the other, and pick the matching synteny track.

<Figure caption="The dotplot import form, with the HG008-T v3.2 assembly on one axis and GRCh38 on the other." src="/img/sv_cgiab/dotplot_import_form.png" />

HG008-T v3.2 is haplotype-resolved, so its scaffold names end in `_hap1` or
`_hap2` and one plot stacks both on the same axis, doubling every diagonal.
Restrict the y axis to one haplotype at a time and each plot reads as a plain
assembly-vs-reference diagonal.

<Figure caption="Haplotype 1 of HG008-T v3.2 (y) against GRCh38 chromosomes (x). Each scaffold is one diagonal segment; scaffolds named for two chromosomes (chr3_chr13_hap1) break into two, which is the translocation." src="/img/sv_cgiab/dotplot_hap1.png" />

<Figure caption="The same plot for haplotype 2. chr13_hap2 carries a single clean diagonal against chr13, the untranslocated counterpart to hap1's fused scaffold." src="/img/sv_cgiab/dotplot_hap2.png" />

Drag over a region and take **Launch → Linear synteny view** from the selection,
keeping **HG008T v3.2** as the dialog's synteny dataset, then enter `chr3 chr13`
in the GRCh38 search box to focus on those chromosomes. Raising the **minimum
alignment length** (in the synteny view's menu) drops short, noisy anchors so
the large syntenic blocks read clearly, and zooming in on a breakpoint reads it
at base level.

<Figure caption="A synteny view launched from the chr3/chr13 selection in the dotplot: GRCh38 chr3 and chr13 above, the fused chr3_chr13_hap1 scaffold and chr13_hap2 below, at a raised minimum alignment length." src="/img/sv_cgiab/synteny_view.png" />

The chr3/chr13 fusion is one of this genome's truncal interchromosomal
rearrangements, many of which break in or near a centromere
([Wagner et al. 2026](https://doi.org/10.64898/2026.05.01.722316)). A scaffold
named for two GRCh38 chromosomes is the cue, so the names on the dotplot's y
axis are a worklist.

For more on these views, see the
[dotplot view guide](https://jbrowse.org/jb2-staging/docs/user_guides/dotplot_view) and the
[linear synteny view guide](https://jbrowse.org/jb2-staging/docs/user_guides/linear_synteny_view).

### Methylation on the tumor reads

The C-GIAB PacBio HiFi BAMs carry per-read 5mC calls in their `MM`/`ML` tags,
and JBrowse renders those with no extra files: open the tumor reads and set
**Color by... → Modifications** from the track menu. The tags survive the
conversion to CRAM above, so the reads loaded for the SV walkthroughs already
carry them.

Two modes sit under that item:

- **One color per modification type** marks the cytosines the basecaller called
  modified and leaves the rest blank, so an unmethylated stretch and a stretch
  with no CpGs in it look alike.
- **One color per type, plus low-probability & unmodified in blue** paints every
  CpG in context, which is what the figure below is set to.

<Figure caption="Tumor PacBio HiFi reads at the CDKN2B-AS1 end of the CDKN2A locus, over the NCBI RefSeq gene lane, colored by base modification with unmodified cytosines filled in. Neighboring CpG-dense blocks come out in opposite states, one of them at the CDKN2B-AS1 transcription start." src="/img/sv_cgiab/methylation_cdkn2b.png" />

Where the marks thin out to scattered ticks, that is CpG density: the fill draws
a cytosine only where the reference puts one in context.

Most somatic LINE insertions in HG008 come from two hypomethylated non-reference
germline LINE insertions, so the methylation state of a source element explains
the insertion burden downstream of it
([Wagner et al. 2026](https://doi.org/10.64898/2026.05.01.722316)). See
[Modifications and methylation](https://jbrowse.org/jb2-staging/docs/user_guides/alignments_track#modifications-and-methylation)
for the display modes, and the
[methylation tutorial](https://jbrowse.org/jb2-staging/docs/tutorials/methylation) for the aggregate and
allele-specific views.

## Where to go next

Swap the VCF, the CRAMs, the caller output and the assembly for your own, and
the same tracks and walkthroughs apply. The
[SV visualization guide](https://jbrowse.org/jb2-staging/docs/user_guides/sv_visualization) covers the display
options the walkthroughs reach for: the color schemes (pair orientation, insert
size), the read filters (discordant pairs, soft-clipped), and the display modes
(pileup, read arcs, linked reads).

Within C-GIAB itself there is more on the same FTP than this tutorial loads:

- a **somatic small-variant draft benchmark**, published alongside the SV/CNV
  one used here, which loads as a variant track the same way
- the **matched normal assembly** (`HG008N`), which loads as a second JBrowse
  assembly and serves as the synteny target. Comparing the tumor assembly to the
  donor's own normal separates somatic change from germline difference, and it
  is the approach the C-GIAB assembly paper is built on
- **HG009**, a second matched pair (PDAC liver metastasis with matched CD4+ T
  cells) on the
  [NIST C-GIAB page](https://www.nist.gov/programs-projects/cancer-genome-bottle)

## Reproduce it end to end

[`build_sv_visualization_cgiab.sh`](https://github.com/GMOD/jbrowse-components/blob/main/scripts/build_sv_visualization_cgiab.sh)
runs the whole data-preparation pipeline above in one shot:

```bash
curl -fO https://raw.githubusercontent.com/GMOD/jbrowse-components/main/scripts/build_sv_visualization_cgiab.sh
bash build_sv_visualization_cgiab.sh   # builds ./cgiab_build/jbrowse2
npx --yes serve cgiab_build/jbrowse2
```

It downloads the C-GIAB GRCh38 build and the V0.5 benchmark calls, turns both
HiFi BAMs into CRAMs, computes megadepth coverage, calls copy number with
HiFiCNV, builds the BAF bigWig, aligns the T2T tumor assembly to GRCh38 with
minimap2, downloads JBrowse, and writes a `config.json` with all of it loaded
beside the published Wakhan segments.

The BAF and Wakhan tracks go in with `add-track-json`: the settings that make
them readable (`resolutionMultiplier` on one, `partitionField` on the other) are
track config, with no command-line flag.

It needs the tools listed under [Prerequisites](#prerequisites), plus `bcftools`
and `bedGraphToBigWig`. It pulls down more than 200 GB, wants roughly 1.5 TB of
free disk and 32 GB of RAM, and its alignment and copy-number steps take hours.

## See also

- [](https://jbrowse.org/jb2-staging/docs/tutorials/synteny_visualization)
- [](https://jbrowse.org/jb2-staging/docs/tutorials/sv_callset_review)
- [](https://jbrowse.org/jb2-staging/docs/tutorials/cancer_sv)
- [](https://jbrowse.org/jb2-staging/docs/user_guides/sv_visualization)
- [](https://jbrowse.org/jb2-staging/docs/user_guides/sv_inspector_view)
- [](https://jbrowse.org/jb2-staging/docs/user_guides/multiquantitative_track)

## References

- Bailey et al. (2016).
  [Genomic analyses identify molecular subtypes of pancreatic cancer](https://doi.org/10.1038/nature16965)
- Diesh et al. (2023).
  [JBrowse 2: A Modular Genome Browser with Views of Synteny and Structural Variation](https://doi.org/10.1186/s13059-023-02914-z)
- McDaniel et al. (2025).
  [Development and Extensive Sequencing of a Broadly-Consented Genome in a Bottle Matched Tumor-Normal Pair](https://doi.org/10.1038/s41597-025-05438-2)
- Rautiainen et al. (2023).
  [Verkko: telomere-to-telomere assembly of diploid chromosomes](https://doi.org/10.1038/s41587-023-01662-6)
- Waddell et al. (2015).
  [Whole genomes redefine the mutational landscape of pancreatic cancer](https://doi.org/10.1038/nature14169)
- Wagner et al. (2026).
  [A complete human pancreatic cancer genome](https://doi.org/10.64898/2026.05.01.722316)