modules/
custom_orfmerge

Description

Cluster normalised per-sample, per-caller ORF predictions into a single cohort-level catalogue. Pair with custom/orfnormalise upstream and (typically) bedtools/getfasta + seqkit/translate downstream to obtain the AA FASTA.

Rows sharing an identical exon structure are grouped first and clustered as a single proxy chosen by class specificity, then the members are restored. Class selects the strategy, so without this an ORF that two callers class differently would be split across strategies and could never merge.

Proxies are then partitioned by clustering strategy. The harmonised orf_class written by custom/orfnormalise selects the strategy but is never part of a grouping key, because callers disagree on class for the same ORF (Ribo-TISH reports 5'UTR for both uORFs and CDS-overlapping uORFs) and keying on it would emit one catalogue row per disagreeing caller:

• canonical_cds: grouped by (transcript_id, strand), then reciprocal-overlap clustered within the transcript so a short truncated variant is not folded into the full-length CDS.
• uORF, uoORF, dORF, collapse by (transcript_id, strand, start, doORF, intORF, other: end). A single transcript can host multiple distinct uORFs / dORFs / internal ORFs, so keying on the outer span keeps them in separate clusters while still merging cross-caller calls that agree on coordinates.
• novel_u: greedy reciprocal-overlap clustering on summed exon-block intersection at --reciprocal-overlap (default 0.8). Catches fuzzy cross-caller matches and exact-coordinate collapses in one pass. Order-dependent at the boundary: a chain A-B-C where A-B and B-C overlap at ~0.85 but A-C only at ~0.75 may cluster as {A,B,C} or {A,B}+{C} depending on iteration order. Rare in practice at 0.8.

Cross-caller consensus is recorded in two column families on the catalogue TSV:

• called_by_<caller>: 0/1 indicator per supported caller (ribotish, ribocode, ribotricer, rpbp, price).
• score_<caller>: best score from that caller within the cluster. Score direction is per-caller (p-values are minimised; Bayes factors / phase scores are maximised).

Cross-sample recurrence is recorded in two further columns:

• n_samples: number of distinct samples contributing to the cluster (a cohort recurrence metric).
• samples: sorted, comma-separated list of those sample ids.

Emits a small MultiQC custom-content TSV (per-class counts) for inclusion in downstream MultiQC reports.

Alongside the full catalogue, emits a consensus view (*.consensus.*) filtered to ORFs supported by at least --min-callers distinct callers and recurring in at least --min-samples samples (both default 1, i.e. no filtering, so the consensus view equals the full catalogue). Raising either threshold yields a higher-confidence catalogue without altering the full one.