intronIC will automatically generate a set of output files. A brief description of the contents of each file follows (numbered lists represent the columns within each file).

Note: The meta.iic and score_info.iic files include column headers by default. Use --no-headers to omit headers if needed for compatibility with legacy pipelines. All .iic data tables are tab-delimited.

Intron label format: the intron name/label (first column of most files) has the form {SpAbbr}-{gene}@{transcript}_{index}({family_size}){tags} — e.g. HomSap-ENSG00000141837@ENST00000614285_1(47). HomSap is a 3+3-character species abbreviation (first three letters of genus + species); the gene and transcript IDs come from the annotation; _1 is the intron's ordinal within the transcript; (47) is the number of introns in that transcript. Optional trailing ;[…] tags encode special conditions (for example [c:-1], a corrected boundary).

For many use-cases, the most important files will be:

• score_info.iic - the authoritative per-intron feature+call matrix (the call column is type_id; the motif probability is P_motif)

• bed.iic - BED file of intron coordinates, with the adjudicated calling score (adjusted_score, 0–100) in the 5th column

• meta.iic - metadata file with intron info such as parent transcript/gene, ordinal index, phase, CDS-relative position, etc.

• introns.iic - full intron sequences, with the (first-pass) SVM score in the 2nd column

A mapping table of unique, scored intron labels and their corresponding duplicate intron labels:

ENST00000359596:intron_38433875_38440744	ENST00000355481:intron_38433875_38440744

1. scored intron key (<transcript>:intron_<start>_<stop>)
2. duplicate intron key sharing those coordinates

A mapping table in the same two-column format as dupe_map.iic (no header), pairing each scored intron key with an overlapping intron key that was excluded in its favour — one row per excluded member. Written only when overlapping introns are found (normally absent).

1. representative (retained) intron key
2. excluded overlapping intron key

All of the annotated intron sequences, including any introns not meeting scoring criteria (e.g. too short, including non-ATCG characters in scoring regions, etc.; includes duplicate introns if run with -d):

HomSap-ENSG00000141837@ENST00000614285_1(47);[c:-1]	100.0	AGCGAAGACAACGTGGTGAGAAAATACGCCAAAAAGATCACCGAATGGCC	ATATCCTTTTGCCCGAACCCCAGCAGCAGCTGCGCCTCCC[...]TCTTTTCTTTGAACAAGATTTTTCCTTAATTCCCCAATAC	TCCCTTTGAATATATGATTTTAGCCACCATCATAGCGAATTGCATCGTCC

(the intron sequence is shown truncated as [...]; real rows carry the full sequence)

1. label (without score tag)
2. score: the first-pass SVM U12-type probability (0–100), or NA if unscored. This is not the adjudicated call; use type_id in score_info.iic/meta.iic for the final U12-type/U2-type label
3. upstream (5′ exon) sequence (50 nt by default, set with --flank-len)
4. intron sequence
5. downstream (3′ exon) sequence (50 nt by default, set with --flank-len)

This file (name, score, upstream flank, intron seq, downstream flank) is also the input format for -q/--sequence-file re-scoring mode.

A BED format file of intron coordinates with the adjudicated calling score and labels:

19	13455212	13505931	HomSap-ENSG00000141837@ENST00000614285_1(47);[c:-1]	100	-	corrected

1. genomic region (e.g. 'chr1')
2. start coordinate (0-indexed)
3. end coordinate (1-indexed)
4. label (with any coordinate-correction tag, e.g. [c:-1]; no score tag)
5. adjusted_score: the adjudicated 0–100 calling scale (synced from score_info.iic after adjudication); ≥ 50 ⟺ u12. NA if unscored
6. strand
7. attributes (comma-separated flag list, same as meta.iic column 16)

Note: The BED file is not created when using sequence-only input (-q flag), as BED format requires real genomic coordinates.

A human-readable run log (<name>.iic.log) of the information generated during operation, including total number of introns processed, excluded, etc., the per-species adjudication summary, and total number of U12-type introns identified. This file is not machine-parseable; use the JSON/TSV outputs for data.

Additional metadata for each intron (ordered by increasing U12-type score):

HomSap-ENSG00000141837@ENST00000614285_1(47);[c:-1]	10.0000	AT-AC	GCC|ATATCCTTTT...TTTTCCTTAATT/TTTTCCTTAATT...AATAC|TCC	CACCTCCAACACCCTTCTTTTCTTTGAACAAGAT[TTTTCCTTAATT]CCCAATAC	-12	50719	ENST00000614285	ENSG00000141837	1	47	0.039	2	u12	cds	corrected

1. label
2. rel_score: signed, centered on the 90% high-confidence threshold, range ≈ [−90, +10]; > 0 marks a high-confidence U12-type call. rel_score < 0 means "below the 90% threshold", not "U2-type". Synced from score_info.iic
3. terminal dinucleotides (e.g. GT-AG, AT-AC)
4. motif_schematic: motif string (5′, U12/U2 BPS, 3′)
5. bp_context: BPS location in the context of the 3′ end
6. bp_offset: branch point adenosine position relative to 3'SS (e.g. -13)
7. length (bp)
8. parent transcript
9. parent gene
10. ordinal position in transcript
11. total introns in transcript
12. frac_pos: CDS-relative position = (CDS length upstream of the intron) / (total CDS length), in [0, 1] (e.g. 0.039 for an intron near the 5′ end of the coding sequence, 0.5 at the midpoint). Exon-only (UTR) introns have no CDS position and are NA
13. phase (0, between codons; 1, after the first base of a codon; 2, after the second base of a codon)
14. type_id: binary classification (u12 or u2), synced from score_info.iic (the authoritative call). A u12 label may include introns with probabilities well below the high-confidence threshold; filter on rel_score > 0 for high-confidence calls only
15. genomic feature used to define the intron (e.g. cds, exon)
16. attributes (comma-separated tags indicating special conditions, e.g. noncanonical, corrected, not_longest_isoform, duplicate)

The authoritative per-intron feature + call matrix: 35 tab-separated columns with a header row (omit with --no-headers). It contains only scored introns (omitted introns appear in meta.iic/bed.iic/introns.iic, not here; join on the name column). Rows are written in two stages: the base writer emits columns 1–26, then the species adjudicator re-reads the file, appends columns 27–35, and overwrites rel_score (col 2) and adjusted_score (col 25). Two missing-value sentinels are used: base columns write NA; the appended adjudicator columns write lowercase nan.

Columns marked [S] are per-species constants (identical on every row); all others are per-intron.

Reading the call: key off P_motif (27) + motif_category (33) + type_id (35). type_id = u12 iff P_motif ≥ 0.5 and motif_category ≠ NOT_DETECTED. svm_score (3) is a different calibration than P_motif and is not the call; don't threshold it.

The columns q, P_adj, P_adj_lo and P_adj_hi were removed in the 2026-07 schema (38 → 35 columns). All four were deterministic functions of P_motif + motif_category: q was a per-species constant, and P_adj_lo/P_adj_hi were bit-identical to P_adj. adjusted_score is now computed directly as 100·q_eff·P_motif.

bg_fdr (col 34) — report-only background surprisal. For each strong call (P_motif ≥ 0.90), the fraction of the calls at least that strong which the genome's own U2 tail already accounts for. It is the per-intron decomposition of z_excess, computed from the same fitted tail, with a Benjamini-Hochberg step-up so it is monotone in call strength. Read ~1 as "indistinguishable from this genome's background" and ≪1 as "the background cannot account for it". It is nan for non-calls (those are the background — rank them with P_motif) and whenever the U2 tail cannot be referenced.

bg_fdr is populated even when the calling columns are zeroed, which is its main use: it is the only column that distinguishes a NOT_DETECTED genome whose calls are uniformly background from one that holds a genuine outlier. In Symbiodinium natans all 109 strong calls sit at bg_fdr = 1 (its best call is weaker than the maximum its own background is expected to produce), while Phytophthora infestans — also NOT_DETECTED, also fully zeroed — holds one call at 6.9e-3. Neither z_excess nor P_motif separates those two: P_motif saturates, reading 1.0000 at the top of both. Within a DETECTED genome it identifies which calls carry the signal.

bg_fdr is report-only per intron — no intron's own call depends on its own bg_fdr. Since v3.1 the minimum over a genome's calls does feed one per-species decision: the low-k escape hatch, which spares a genome with only one or two strong calls from being zeroed when the background cannot explain the strongest of them.

Per-species aggregate summary written alongside each output (27 keys, all always present; adjudicator statistics become JSON null when the genome's U2-type tail can't be referenced):

Totals nest: total_introns_generated ⊋ total_introns ⊋ total_scored; u12_count + u2_count == total_scored; high_confidence_u12 ⊆ u12_count.

The substrate_quality object is a report-only confidence/QC layer — it annotates a genome and flags annotation problems, but never changes a U12/U2 call. Read it together with motif_category and the snRNA evidence: a low signal from a polluted annotation (a real bearer whose gene models carry junk introns) is a different thing from a degenerate splicing signal (a divergent, snRNA-absent genome).

Typical reads: annotation_quality = poor/suspect on a confirmed bearer flags the assembly/annotation for review (e.g. polyploid or repeat-rich genomes with over-predicted gene models), not the biology; splicing_signal = degenerate on an snRNA-absent genome marks the divergent / ambiguous detections (e.g. dinoflagellates). An all-unassessed block just means the genome had too small a U2-type background to assess.

A per-species adjudicator sidecar (emitted by the pmotif_adjudicated bundle) holding the fitted U2-tail model and the numbers behind motif_category. Always-present keys include params_version, platt_a/platt_c, call_threshold (0.90), u2_threshold (0.50), n_u2, n_call, z_excess, cs_p95, p_gumbel_p95, motif_category, call_margins (the per-call ensemble margins), assessable/reason, and the EVT tail-selection parameters evt_tail_pct (90.0) and evt_tail_frac (0.10). When the genome has ≥ 200 U2-type introns it also carries the fitted EVT tail (med, mad, q90, lam, exp_max, q_tail, plus a pre-binned U2-type histogram). This sidecar drives the per-species tail-model diagnostic plot (.plot.tail_model.iic.png), which shows the U2-type background margins, the fitted U2-type tail extrapolated into the call region, and the U12-type calls that escape it.

intronIC renders a set of per-species diagnostic plots (PNG) in background-corrected raw-motif-log-odds space. Two are always produced; the rest are conditional.

(Plots suffixed .AUC.iic.png, .ref_hex.iic.png, .plot.training_scatter.iic.png, and .plot.decision_surface.iic.png are produced only when training a model, never by a classification run.)

The species-agnostic, Platt-calibrated probability that an intron's motif is U12-type, P_motif = σ(2.796·margin − 1.178) on the SVM ensemble margin. It does not depend on how many U12-type introns the genome has, so you can threshold it post-hoc at whatever confidence you need:

svm_score (column 3 of score_info.iic) is a different isotonic calibration retained for auditability; it is not the call — use P_motif / type_id.

adjusted_score = 100·q_eff·P_motif, where q_eff ∈ {0,1} is the binary effective gate (q = 0 only when the genome's motif_category is NOT_DETECTED, else q = 1). So in any non-NOT_DETECTED genome adjusted_score = 100·P_motif; in a NOT_DETECTED genome all calls are zeroed. The call is adjusted_score ≥ 50 ⟺ u12. This is the adjudicated calling scale.

rel_score = adjusted_score − 90 (range ≈ [−90, +10]). It is signed and centered on the 90% high-confidence threshold, not on the U12-type/U2-type boundary: rel_score > 0 ⇔ adjusted_score > 90 ⇔ high-confidence U12-type. rel_score < 0 means "below the 90% threshold", not "anti-U12-type" — a U2-type call and a sub-high-confidence U12-type call are both negative.

Each motif region's raw score is a background-corrected log-odds ratio log₂(P(seq | U12) / P(seq | U2)), with a different natural range per region. v3 scores these raw log-odds directly — there is no per-species z-normalization (5'_raw, bp_raw, 3'_raw in score_info.iic). See Technical Details for the scoring approach.

The authoritative per-intron call is type_id (score_info.iic col 35, or meta.iic col 14). Filtering on rel_score > 0 selects the high-confidence U12-type subset.

# All U12-type calls, from meta.iic (col 14 = type_id)
awk -F'\t' '$14=="u12"' species.meta.iic

# High-confidence U12-type introns, from meta.iic (col 2 = rel_score, > 0 ⇔ adjusted_score > 90)
awk -F'\t' '($2!="NA" && $2>0)' species.meta.iic

# From bed.iic (col 5 = adjusted_score, the adjudicated calling scale)
awk '$5 >= 50' species.bed.iic

# Count high-confidence U12-type introns
awk -F'\t' '($2!="NA" && $2>0)' species.meta.iic | wc -l

# AT-AC introns only
awk '($3 == "AT-AC")' species.meta.iic

# High-confidence U12-type AT-AC introns
awk '($2 > 5 && $3 == "AT-AC")' species.meta.iic

# U12-type introns in first half of transcript
awk '($2>0 && $12 < 50)' species.meta.iic

# Intron sequences with a strong first-pass score (col 2 = svm_score, a rough proxy —
# for the authoritative call, filter meta.iic on type_id and join on the label).
awk -F'\t' '$2!="NA" && $2 > 90 {print ">"$1"\n"$4}' species.introns.iic > u12.fasta