Fix #394: in-frame stop-codon deletion with empty 3' UTR is a Deletion

[iskandr]

Problem

Closes #394.

An in-frame deletion that removes the stop codon of a transcript with no 3' UTR sequence crashes:

from pyensembl import ensembl_grch37
import varcode
v = varcode.Variant(contig='16', start=30128151, ref='GGGATGCCTACGTGCCCCC', alt='G', ensembl=ensembl_grch37)
v.effects()
# ValueError: If no amino acids added by StopLoss then it should be Silent

(The reporter's original start=30128150 is off-by-one against the GRCh37 reference; 30128151 is the equivalent ref-matching coordinate. Same MAPK3 region, same crash.)

Root cause

translate_in_frame_mutation sets using_three_prime_utr=True whenever the mutation runs past the reference stop codon — even when the transcript has no 3' UTR sequence to translate into (transcript.three_prime_utr_sequence == "", e.g. MAPK3-006 / ENST00000395199). predict_in_frame_coding_effect then takes the StopLoss branch with an empty aa_alt, which the StopLoss constructor rejects.

Fix

Only emit StopLoss when readthrough actually adds residues (n_aa_alt > 0); otherwise fall through to the existing n_aa_alt == 0 branch, which reports a C-terminal Deletion. With no UTR sequence we cannot predict an extended protein, so a Deletion is the honest classification — matching the suggestion in #394.

The earlier #246 fix only handled transcripts with a non-empty 3' UTR.

-    elif using_three_prime_utr:
+    elif using_three_prime_utr and n_aa_alt > 0:
         return StopLoss(...)
     elif n_aa_alt == 0:
         return Deletion(...)

Tests

The reported MAPK3 variant conveniently exercises both sides of the fix from a single input:

• ENST00000395199 (no 3' UTR) → Deletion p.GGT354del — the fixed crash
• ENST00000403394 (804 nt 3' UTR) → StopLoss readthrough — proves the Deletion fallback wasn't over-broadened
• v.effects() across all MAPK3 transcripts no longer raises
• plus a splice-free unit test of predict_in_frame_coding_effect for the empty-UTR branch

Added to tests/test_stop_codon_classification_bugs.py. Full suite passes (the one unrelated test_filter_effects_by_transcript_expression failure is pre-existing on main — it's sensitive to the locally-installed Ensembl release).

https://claude.ai/code/session_0149VWj5Rm1rYFf9azu4ry62