2.17.0 — a truncated junction can be finished from the germline

[mikessh]

--complete-junctions N (off by default)

A read that reaches Cys104, runs into the J and stops before [FW]118 has no observed 3' boundary, so arda declined it. The J's 5' chew-back and the N/P additions all lie upstream of the read's last aligned J base, so what is missing is germline-templated — and the reference already stores exactly that span (a J's germline_nt runs the allele's 5' end through the [FW]118 codon), so the tail is germline_nt[j_germline_end:] and nothing has to be aligned to find it.

⛔ There is no V-side counterpart and there will not be one: a read short at the 5' end is missing bases the V germline does not template either, which is what v_anchor_prefix refuses.

library	junctions	completed	median nt imputed
bulk RNA-seq, 1 M pairs (SRR5233639)	3,856 → 4,103 (+6.4 %)	247	13
TRA amplicon, 100 k reads	44,497 → 44,527 (+0.07 %)	30	12

The bases are imputed, so every completed row carries the count in junction_completed_nt (appended last, after the anchor columns) rather than leaving a consumer to infer it. No already-observed junction moves in either arm.

The refusal is most of the feature. A read whose alignment stops more than a partial codon short of its own 3' end is declined — on the TRA amplicon 236 of 266 candidates run from the V straight into TRAC with no J at all, while the aligner still names a J off a few coincidental bases. Of the 30 that survive: 30/30 close on [FW]118, 22 of 23 extend the (itself truncated) IgBLAST junction, and 7 are byte-identical to a junction the same library observes in full on other reads.

⚠ On IG the imputed span can hide the SHM the read would have shown, biasing a completed junction's 3' end toward germline — and IG is where the yield is (175 of 247 bulk completions are IGH).

_segment_rows refuses a target-inverted row, as _best_hits already did

_SEGMENT_FORMAT never asked MMseqs2 for tend, so 2.10.0's guard covered one of the two reductions and not the other. A segment row's tstart is read as a forward target offset by both consumers; _cannot_reach_cys104 walks it forward, so an inverted row can answer "unreachable" for a read that does reach Cys104 — which is then aligned V-segment-only and emits no junction. Measured cost of the extra column: none (same-job alternating A/B, master 9.78–10.08 s vs 9.64–10.05 s).

Measured, not changed

--call-level gene shipped in 2.16.0 with its polyclonal cost unmeasured. Now: IGH −4.39 %, IGK −2.58 %, IGL −3.16 %, TRB 0.00 %, TRA −0.13 %, reads conserved exactly on both arms.

Tests

All three tiers green (738 unit + 62 synthetic/realworld). Every new test verified to fail without its fix.

🤖 Generated with Claude Code