fix(hgvs): route in-frame insertions to delins, not substitution

[kuanlinhuang]

Fixes the half of #81 that produces a silent wrong answer. Protein-level normalisation (3'-shift and dup collapse) is not in this PR and is left to @MarcusOlivecrona, who has an implementation in flight.

The bug

The HGVSp routing tested aa.1 == "-" || consequences.contains(InframeDeletion). An in-frame insertion satisfies neither, so it fell through to hgvsp(), which compares one residue per side:

if ref_aa == alt_aa {
    return Some(format!("{}{}{}=", prefix, ref_aa3, protein_pos));
}

For amino_acids R/RR that yields p.Arg3=, a synonymous call for a variant that lengthens the protein. For E/DVDFREYE it yields p.Glu598Asp, a well-formed missense. Unlike the malformed p.Xaa123??? that #58 fixed, both are indistinguishable from genuine calls, so nothing downstream can detect them. In the reporter's clinical panel run all 15 distinct in-frame insertions were affected (8 synonymous, 7 missense), two of them FLT3 ITDs.

Reproduced through the real annotation path before fixing:

consequence_terms = ["inframe_insertion"]
amino_acids       = "R/RR"
codons            = "cgg/CGGcgg"
hgvsp             = "ENSP_INS_TEST:p.Arg3="     <- Ensembl VEP: p.Arg3dup

The change

In-frame insertions now route into hgvsp_inframe_deletion alongside deletions and delins, giving p.Arg3delinsArgArg.

That output is deliberately un-normalised. Ensembl VEP collapses this repeat to p.Arg3dup, and for a 31-codon BCOR duplication the delins form runs to ~105 characters. That is the tradeoff being accepted here: correct-but-verbose now, normalised when #81 lands. What matters clinically is the property this PR establishes, that an in-frame insertion never renders as a substitution.

Both call sites are fixed. fastvep-annotate (library/web) and fastvep-cli (CLI pipeline) carry separate copies of this routing block, and both had the bug. Nothing structurally keeps them in step, so each gets its own regression test.

Testing

• crates/fastvep-annotate/src/lib.rs - annotates a VCF against a synthetic transcript through annotate_vcf_text_with_acmg, asserting the emitted HGVSp.
• crates/fastvep-cli/tests/hgvsp_inframe_insertion.rs - end-to-end through run_annotate with a GFF3 + FASTA + VCF on disk, asserting on the CSQ field in the output VCF.

Both fail on master with p.Arg3= and pass here. Both assert the inframe_insertion consequence first, so they cannot pass for the wrong reason if classification changes.

cargo test --workspace (0 failures) and cargo clippy --workspace --all-targets (clean) both pass.

cargo fmt --check is not clean on master and is not made worse here: the violation count in the two touched files is 111 before and after this change, and the new test file is rustfmt-clean on its own. The pending chore/rustfmt-the-workspace branch reformats the workspace separately and will absorb these lines.

ACMG is untouched by construction

We just shipped a large ACMG pass (#82), so the blast radius matters. ACMG never reads hgvsp - PS1, PM1 and PM5 read protein_position and amino_acids directly, and the PVS1 truncation percentage reads protein_position. This change touches neither, so no classification can move. That containment is also the constraint requested of the normalisation work in #81.

Refs #81

🤖 Generated with Claude Code