`manta/germline`: pinned container segfaults in `GenerateSVCandidates` on WGS trios

Description of the bug

The container pinned by manta/germline segfaults reproducibly part-way through
GenerateSVCandidates on a whole-genome trio. Illumina's official Manta 1.6.0
release build completes the identical workload. The difference is the build, not
the data.

Discovered through nf-core/raredisease 3.1.2, whose CALL_SV_MANTA subworkflow
runs this module. raredisease cannot skip Manta alone — skip_tools has no
manta entry and SVDB_MERGE requires one priority tag per VCF — so the only
available workaround is disabling the whole sv_calling subworkflow.

Container pinned

community.wave.seqera.io/library/manta_python:0eb71149179b3920

Steps to reproduce

Human WGS trio, GRCh38 (hg38.analysisSet.fa), ~30x per sample, aligned with
bwa-mem2 and duplicate-marked (stock nf-core/raredisease alignment; no ABRA2 or
external realignment). Joint three-sample invocation with the pinned
Clinical-Genomics call regions:

configManta.py --bam A.bam --bam B.bam --bam C.bam \
  --reference hg38.analysisSet.fa --runDir manta \
  --callRegions grch38_manta_call_regions_-1.0-.bed.gz
python manta/runWorkflow.py -m local -j 12

Observed

generateCandidateSV_0000 dies with SIGSEGV and no stderr:

[generateCandidateSV_0000] [wrapperSignal] taskExitCode -11
Last 0 stderr lines from task (of 0 total lines):

Four runs, three thread counts, same wall:

threads	edges processed	result
6	4806	SIGSEGV
12	4820	SIGSEGV
12	4819	SIGSEGV (enableRemoteReadRetrieval...=0)
1	4994	SIGSEGV (deterministic order, so not a race)

Expected

Completion, as with Illumina's official release build on identical inputs:

EXIT=0   368 tasks, 0 errors, 9h11m
results/variants/diploidSV.vcf.gz          18,789 records
results/variants/candidateSV.vcf.gz        7.1 MB
results/variants/candidateSmallIndels.vcf.gz  4.5 MB

Same BAMs, same reference, same call regions, same thread count, same host.

Ruled out by experiment

• Memory / disk. Peak RSS 1.7 GB against a 36 GB request and 3.2 GB against
  72 GB. SLURM reports FAILED 1:0, not an OOM kill.
• Thread race. Reproduces single-threaded at a comparable edge count.
• enableRemoteReadRetrievalForInsertionsInGermlineCallingModes. Disabling
  it changes nothing.
• chrM. chromDepth.txt reports chrM at 48697x against ~100x autosomal, and
  Manta's depthFilterMultiple guard is relative to each chromosome's own
  median so it never engages there — but a chrM-free call-regions BED crashed
  identically (4819 edges).
• The locus graph. SummarizeSVLoci output is byte-identical between the
  two builds. Graph construction is correct in both; only candidate generation
  differs. (The two svLocusGraph.bin files differ in bytes and the official
  binary rejects the container's with FATAL_ERROR: unsupported version, but
  that is a Boost archive-format difference, not a semantic one.)
• Input integrity. All BAMs pass samtools quickcheck; @PG chain is
  bwa-mem2 -> samtools -> MarkDuplicates -> samtools.

Suspected cause

Manta 1.6.0 vendors boost_1_58_0 in redist/ and is built and tested against
it. The two builds:

	Pinned container	Illumina official release
GenerateSVCandidates	10,310,880 bytes	7,621,512 bytes
Boost	libboost_serialization.so.1.85.0, dynamic	static, vendored 1.58
Compiler	g++-13.3.0	g++-6.3.1
Build date	2025-02-04	2019-06-28

Manta has been unmaintained since 2019, so a 2025 rebuild of that C++ against a
Boost 27 minor versions newer is a plausible route from latent undefined
behaviour to a segfault. This was not isolated — compiler and htslib differ
too — but the remedy is the same regardless of which difference is responsible.

Suggested fix

quay.io/biocontainers/manta:1.6.0--h9ee0642_1 (2021-03-26) is the same build
shape as the official release: static Boost, GenerateSVCandidates 7,625,608
bytes against the official 7,621,512. Available as a Singularity image at
https://depot.galaxyproject.org/singularity/manta:1.6.0--h9ee0642_1.

The trade-off is an older htslib in that image. An alternative is a fresh
bioconda build pinned to Boost 1.58, which would need a bioconda-recipes
change.

Caveat for whoever triages this

This will most likely not reproduce with -profile test. The trigger looks
data-dependent: this trio's locus graph contains a pathological hub node at
chr2:32,915,715-32,917,079 (1,364 bp) with 39,924 edges against a
locus-wide median of 1 and p99 of 17, and 1,502,934 inbound observations —
nearly every node in an 84,061-node locus points at it. Subsampled test data is
very unlikely to build such a node. Reproducing this needs a real WGS trio.

Happy to supply the locus graph, SummarizeSVLoci output, or the full pyflow
logs on request.

Command used and terminal output

Through nf-core/raredisease.

samplesheet.csv:

sample,lane,fastq_1,fastq_2,sex,phenotype,paternal_id,maternal_id,case_id
NA10837,L001,ERR3989284_1.fastq.gz,ERR3989284_2.fastq.gz,1,2,NA12272,NA12273,G1K-trio
NA12272,L001,ERR3239465_1.fastq.gz,ERR3239465_2.fastq.gz,1,1,,,G1K-trio
NA12273,L001,ERR3239466_1.fastq.gz,ERR3239466_2.fastq.gz,2,1,,,G1K-trio

params.yaml (resource paths elided):

analysis_type: wgs
platform: illumina
genome: GRCh38
aligner: bwamem2
variant_caller: deepvariant
mito_name: chrM
skip_tools: gens,germlinecnvcaller
skip_subworkflows: snv_annotation,sv_annotation,mt_annotation,mt_subsample,me_calling,me_annotation,generate_clinical_set

nextflow run nf-core/raredisease -r 3.1.2 -profile singularity \
  --input samplesheet.csv --outdir results \
  -params-file params.yaml

Relevant files

1000 Genomes 30x on GRCh38 (NYGC), CEPH family 1418 trio. Full FASTQs, not
subsampled — the on-disk sizes match ENA's published byte counts exactly.

Sample	Role	Run	FASTQ
NA10837	proband (son)	ERR3989284	_1.fastq.gz · _2.fastq.gz
NA12272	father	ERR3239465	_1.fastq.gz · _2.fastq.gz
NA12273	mother	ERR3239466	_1.fastq.gz · _2.fastq.gz

MD5s (ENA filereport API, result=read_run):

ERR3989284_1  5352a52cfa92cad5523435545088563c   12985369337 bytes
ERR3989284_2  8fe3b0655b7c1efd3ff2eb5e692aa1fa   14116751232 bytes
ERR3239465_1  cbf08408b0b10b6d491f6e235c8c7d80   13228284415 bytes
ERR3239465_2  c2a8eb74b6f34cc5cd8a2e0b67b201db   14233627994 bytes
ERR3239466_1  5c58f63696396dc7e18ea2c0f9a8e4a3   13195731876 bytes
ERR3239466_2  dc7b2757cd30951d2657d5d444c588b0   14284008497 bytes

Roughly 37x per sample, Illumina NovaSeq 6000, 2x150 bp; 82 GB of FASTQ total.
Data collection: https://www.internationalgenome.org/data-portal/data-collection/1000genomes_30x

Reference: UCSC hg38.analysisSet.fa (GCA_000001405.15).
Manta call regions: the pinned Clinical-Genomics artifact raredisease documents,
rare-disease/region/grch38_manta_call_regions_-1.0-.bed.gz at commit
eea53902baa56a05790fec49e256f22cf3454199
(sha256 fc5728b7f0e6fd18aa4a8015db4892cd9ba6cc7c97c6cf93589f2a313939cca5).

System information

No response

Start with the manta/germline module and its pinned community.wave.seqera.io/library/manta_python:0eb71149179b3920 container, then inspect the GenerateSVCandidates entry point and the suggested Bioconda image. Reproduce with the listed full WGS trio and python manta/runWorkflow.py -m local -j 12; done means the workflow completes without SIGSEGV and the module uses a validated build.