The population pattern you describe is the biggest clue, and it points to artifact before you even open IGV: ~100 inversions, every one present in all 8 cases, none in any of the 4 controls, and none homozygous. Real germline inversions don't distribute like that -- you'd expect a range of genotypes (some hom-alt) and sharing that doesn't line up perfectly with the case/control split. A signal that is perfectly case-specific and uniformly heterozygous almost always means a technical difference between the two groups (sequencing batch, library prep, read length, insert-size distribution, aligner/reference version), not biology. LUMPY is very sensitive to the insert-size distribution, so if cases and controls were prepped/sequenced differently that alone can manufacture group-specific "inversions." First check: are the insert-size distributions and read lengths comparable across the two groups?

What a real inversion looks like in IGV:

• Colour alignments by pair orientation and view as pairs. A true inversion gives read pairs in the same orientation (F-F / R-R) instead of normal FR -- IGV colours these distinctly (the non-blue pair colours).

• It has TWO breakpoints, so expect two separate clusters of aberrant pairs (one at each end), plus split / soft-clipped reads whose clipped parts map to the opposite breakpoint.

• Coverage stays roughly normal across the region (an inversion is copy-neutral).

Artifact signatures:

• Aberrant pairs piled at a single spot with no clean second breakpoint, or a diffuse smear rather than two tight clusters.

• Many MAPQ0 / multi-mapping reads, or the locus sits in a repeat, segmental duplication, low-complexity region, or near a reference gap. Turn on colour-by-mapping-quality.

• The same call recurring across samples only in one group (your situation).

Concrete next steps:

1. These calls are un-genotyped -- LUMPY only reports breakpoints. Genotype across ALL 12 samples with smoove/svtyper or duphold; "absent in controls" from a breakpoint-only call often just means under-called, not truly absent.

2. Require concordance with a second caller (Manta or delly). Group-specific calls only one caller makes are usually artifacts.

3. Intersect with a blacklist / segmental-dup / repeat track and drop calls inside them.

4. Confirm insert-size and read-length parity between cases and controls.

If after genotyping a call still shows the clean two-breakpoint pair-orientation signature + split reads and real support in cases but reference-orientation reads in controls, it's worth believing. Right now the uniform case-only-heterozygous pattern says start from the assumption of a batch artifact.