2.19.0 — correct --flag-chimeras

[mikessh]

A template-switch chimera is a query junction explained by two more abundant parents as prefix + suffix across one breakpoint. chimera_parents names them and the breakpoint. Off by default; it never drops a row.

⛔ The germline trap, measured — it is the whole filter

A junction is V 3' tail + N/P/D + J 5' head, and both tails are germline: every clonotype on a V starts with the same bases, every one on a J ends with them. On a TRA amplicon the median V-templated prefix is 10 nt and the median J-templated suffix 25 nt, against a median clone-specific core of 5 nt.

So the same prefix/suffix test run on the raw junction rediscovers the germline and calls 52.20 % of clonotypes chimeric (35.50 % of reads). With the templated tails excluded and 6 non-templated nt required each side: 0.02 %.

⛔ Cell Ranger's rule does not port to bulk, and not because of the constant

Contigs sharing a V prefix ≥ 25 nt with differing CDR3s is a chimera signature within a barcode, where there is ~1 clone per chain. A polyclonal bulk repertoire has thousands of real clones per V gene — there, that rule describes almost every pair. What ports is UCHIME's shape, two more abundant parents, because the abundance ordering rather than the partition is what makes it a claim.

Measured rate

library	clonotypes	flagged	reads
Bulk RNA-seq (IG)	1,741	7 (0.40 %)	13 (0.18 %)
TRA amplicon	19,844	1 (0.01 %)	1

A 20× bulk-vs-amplicon enrichment, in the direction template-switch chemistry predicts — but far too small to justify deleting clonotypes, and the signature cannot separate a true chimera from two real clones that happen to share a prefix and a suffix. Hence a flag.

Requires the reference: with no anchors the germline cannot be excluded, so it emits nothing rather than a germline-driven guess.

Tests

⚠ Every guard is mutation-tested, and three of the first five tests passed with the bug in:

• the germline test needed a query sharing only germline with each parent (the first version had no suffix parent, so nothing fired however wrong the code was);
• the abundance test needed the query to sort last — setdefault on a descending list otherwise makes it claim its own prefix and the check is never consulted;
• the SHM test needed both parents to genuinely match.

A fifth guard proved unreachable — an identical twin has Hamming 0, already caught by the ≤ 2 SHM check — and was deleted rather than left looking load-bearing.

858 pass across three tiers. Default output byte-identical; the column appears only with the flag.

🤖 Generated with Claude Code