-
Notifications
You must be signed in to change notification settings - Fork 1
Glossary
Definitions of the biological, statistical, and software terms used across the intronIC v3.1.0 documentation. Terms are grouped by topic and ordered alphabetically within each group. Each entry is a single concept; where a term has a canonical home elsewhere in the wiki, that page is linked. Works cited by short form (e.g. Moyer et al. 2020) are listed in full under References.
The downstream boundary of an intron, where the spliceosome cleaves and joins the intron to the next exon. It ends in a terminal dinucleotide (AG for the canonical major-spliceosome case, AC for the AT-AC minor subtype) and, in U2-type introns, is preceded by a polypyrimidine tract. intronIC scores a window from −6 to +4 relative to the intron end.
The upstream boundary of an intron, recognized first during spliceosome assembly. Its terminal dinucleotide is GT (or GC) in most introns and AT in the AT-AC minor subtype. The 5′ splice site carries the strongest type-discriminating signal; intronIC scores a window from −3 to +9 relative to the intron start.
The first two and last two intron bases, named 5′-3′ (e.g. GT-AG). GT-AG is by far the most common; GC-AG is a minor major-spliceosome variant; AT-AC is the diagnostic (though not exclusive) terminal dinucleotide of the minor spliceosome. Roughly 25% of U12-type introns are AT-AC and ~75% are GT-AG, so terminal dinucleotide alone does not classify an intron.
The intronic adenosine whose 2′-OH attacks the 5′ splice site to form the lariat intermediate, embedded in a branch point sequence (BPS). The major spliceosome uses a loose consensus 18–40 nt upstream of the 3′ splice site; the minor spliceosome uses a highly conserved motif (consensus TCCTTAAC) typically 10–15 nt from the 3′ splice site. intronIC scans a −55 to −5 window and scores the branch point against 12 bp PWMs derived from CoLa-seq data (Zeng et al. 2022).
Where an intron falls relative to the reading frame: phase 0 lies between two codons, phase 1 after the first base of a codon, phase 2 after the second. U12-type introns are depleted in phase 0 (~30%) relative to the all-intron background (~50%); the phase 1 versus phase 2 split varies by lineage.
Alternative transcript variants of a gene. By default intronIC scores only introns with unique coordinates from the longest annotated isoform of each gene, so an intron shared by several isoforms is counted once and named from the longest isoform (adjustable via -i/-d).
The ribonucleoprotein machine that removes ~99.5% of introns, built from the U1, U2, U4, U5, and U6 snRNPs. It recognizes GT-AG and GC-AG introns with a loose branch point consensus. Introns spliced by it are U2-type.
The machine that removes the ~0.5% of introns that are U12-type, built from the U11, U12, U4atac, U6atac, and (shared) U5 snRNPs. It recognizes both AT-AC and GT-AG introns via a highly conserved 5′ splice site and branch point. Though rare, U12-type introns are ancient, conserved across most eukaryotes, and effectively never gained de novo, so their absence in a lineage reflects loss.
A pyrimidine-rich (C/T) stretch between the branch point and the 3′ splice site, a recognition element for the major spliceosome. It is generally weak or absent in U12-type introns; intronIC records several PPT descriptors (fraction pyrimidine, longest run, T-weighted score) but they are diagnostic, not primary classification features.
Small nuclear RNAs (snRNAs) and the small nuclear ribonucleoproteins (snRNPs) they form are the catalytic and recognition components of the spliceosome. The major spliceosome uses U1/U2/U4/U5/U6; the minor spliceosome substitutes U11 for U1, U12 for U2, and U4atac/U6atac for U4/U6, sharing U5. snRNA gene detection is used only at intronIC calibration time (to corroborate bearer/loss anchors), never by the runtime classifier.
An intron excised by the minor (U12-dependent) spliceosome, distinguished by a strong, conserved 5′ splice site and branch point motif rather than by its terminal dinucleotide. Classifying an intron as U12-type versus U2-type is intronIC's central task.
An intron excised by the major (U2-dependent) spliceosome: the ordinary case for the vast majority of eukaryotic introns. In intronIC output, U2-type is the default (type_id = u2) call for any intron not classified as U12-type.
intronIC declares 6 input motif features (5'_raw, bp_raw, 3'_raw, bp_offset, bp_scan_confidence, support2_raw). Inside each sub-model's sklearn pipeline a transformer deterministically appends 3 engineered interaction terms (corr_5_bp, corr_bp_3, gap_5_bp), so the fitted SVC operates on 9 features (n_features_in_ = 9). The extra three re-express 5′/BP/3′ interactions to shape the RBF kernel; they carry no new measurement.
Blending each species' own nucleotide composition into the U2-type denominator PWMs before scoring, so that raw motif log-odds reflect true splice signal rather than genome-wide base bias. The blend is a Bayesian shrinkage w = n/(n + n0) (default n0 = 1000; see Bayesian shrinkage (n0)). This is intronIC's only per-species feature adaptation; it deflates, rather than inflates, composition-driven false signal.
The distance of the selected branch point adenosine from the 3′ splice site (a negative integer, e.g. −13). It is one of the six declared model features; U12-type introns place the branch point in a tight, characteristic range.
A model feature quantifying how confidently the branch point scan selected its adenosine over competing positions in the search window. Low confidence flags ambiguous branch point placement.
The raw-feature SVM ensemble's signed decision_function output for an intron, its distance from the decision boundary. This single quantity feeds Platt calibration to produce P_motif and (at the genome's upper tail) drives the strength gate. The terms decision_function, decision distance, and margin all refer to this same ensemble margin; intronIC uses the un-clipped margin (not the isotonic probability, which saturates at 0/1).
The standard deviation of an intron's per-model probability across the 42 ensemble sub-models, a per-intron uncertainty descriptor. It is reported for auditability and is not part of the call.
A monotone, non-parametric mapping from a model score to a probability, used internally by each SVM sub-model. The bundle's isotonic probability saturates near 0 and 1, which is why P_motif is derived from the margin via Platt calibration instead. The isotonic-based svm_score column is retained only for auditability.
A motif score of the form log₂( P(seq | U12) / P(seq | U2) ), computed per region (5′ splice site, branch point, 3′ splice site) against the corresponding PWMs. Positive values favor the U12-type model. These background-corrected raw log-odds (5'_raw, bp_raw, 3'_raw) are the raw motif features the classifier consumes.
The species-agnostic, Platt-calibrated probability that an intron's motif is U12-type: P_motif = σ(2.796·margin − 1.178) on the ensemble margin. It does not depend on how many U12-type introns a genome contains, so it is comparable across species and can be thresholded post-hoc at any confidence level (0.5 for a call, 0.9 for high confidence). P_motif is the authoritative motif probability; see Output-files.
Fitting a logistic (sigmoid) function to map an SVM margin onto a calibrated probability (Platt 1999). intronIC's frozen fit is σ(2.796·margin − 1.178) — the sigmoid fit of the ensemble margin against binary U12-type/U2-type labels (leave-clade-out ECE ≈ 7×10⁻⁴). The two constants are preserved across adjudicator gate re-stamps and are not retrained.
A per-position table of the four nucleotide frequencies observed at a motif, used to score how well a candidate sequence matches that motif. intronIC ships separate U12-type and U2-type PWMs for each region and terminal-dinucleotide subtype; scoring an intron produces the raw motif log-odds. See Training-data-and-PWMs.
The background-corrected raw motif log-odds (5'_raw, bp_raw, 3'_raw) plus bp_offset, bp_scan_confidence, and support2_raw — the six features the v3 ensemble consumes directly, without per-species z-normalization. Scoring the absolute (raw) log-odds, rather than z-scored features, is the core v3 architectural change (see z-normalization (removed in v3)).
The classifier is an ensemble of 42 sub-models (210 sub-estimators total), each a support-vector machine with a radial basis function (RBF) kernel (C = 200, γ = 0.001). Bagging (training sub-models on resampled data and averaging them) reduces variance and yields the per-intron ensemble_sigma. The ensemble's averaged margin, not any single model, drives the call.
A model feature (the raw analog of the earlier support2) summarizing corroborating motif support for an intron. It is one of the six declared inputs.
A first-pass, isotonic-calibrated U12-type probability (×100) stored in score_info.iic. It is a different calibration from P_motif and is not the call column; do not threshold it. Use P_motif and type_id. See Output-files.
Per-species rescaling of motif features to a common location/scale, used through v2.x to make scores cross-species comparable. It was removed in v3: leave-clade-out analysis showed it manufactures loss-species false positives (raw-feature AUC 0.916 vs z-normalized 0.786; see AUC). All per-species adaptation now lives in the output-level species adjudicator, and --load-normalizer/--save-normalizer are accepted-but-ignored no-ops.
The two frozen z_excess thresholds that bound the adjudicator gate. loss_ceiling_z = 2.60 is the empirical highest-z_excess snRNA-confirmed U12-loss in the calibration panel (Aspergillus coremiiformis, which retains only U4atac/U6atac); at or below it, calls are consistent with the U2-type background → NOT_DETECTED. bearer_floor_z = 5.50 is a trust threshold (not a separator), widened from 4.00 because the near-floor zone is genuinely mixed (loss Monocercomonoides at z ≈ 4.24 sits adjacent to bearer Blyttiomyces at z ≈ 4.54), so above 5.50 a count-based DETECTED call is trustworthy without corroboration. The [2.60, 5.50) gap is INCONCLUSIVE.
A U12-type bearer — a species whose genome contains a functional minor spliceosome and a population of U12-type introns. The adjudicator's job is to distinguish genuine bearers from U12-loss lineages whose motif-strong false positives merely reflect their U2-type background.
The minimum number of calls (= 3) a genome must have before its upper-tail call-strength percentile is evaluated; below it, no strength gate is applied. Paired with min_u2 as the two power thresholds guarding the per-genome tests.
The 95th percentile of the un-clipped ensemble margins over a genome's calls — a measure of how strong its strongest calls are. cs_p95 ≥ 5.0 is a null-free co-fallback in the strength gate; 5.0 corresponds to a per-genome ~0.5% outlier test (cs_p95 = 5.0 ↔ p_gumbel_p95 ≈ 0.005), retained because it stays stable where a genome's EVT fit is noisy.
The statistical framework for the distribution of maxima of many draws (Coles 2001). intronIC fits, per genome, a Gumbel extreme-value null to its U2-type margin tail, a model of how large a margin the genome's own background noise can produce by chance. Both z_excess and p_gumbel_p95 are computed against this per-genome null, so every species is judged against its own composition.
A motif-strong call in a genome that is a true U12-loss (or U12-absent) lineage, i.e. a U2-type intron whose motif happens to resemble U12-type. Every genome has an irreducible motif-strong floor of these; the adjudicator's anchors are calibrated so that loss-FPs do not trigger a DETECTED species call.
A U2-type training or evaluation intron whose motif closely mimics U12-type, the U2-type introns nearest the decision boundary. Their density at the boundary is what teaches the ensemble a tight U12-type region; the deployed model trains on all positives plus an imbalanced (class-weighted) hard-negative set.
A U12-loss (or U12-absent) lineage — one that has lost the minor spliceosome (and hence its U12-type introns), documented in C. elegans, various fungi (e.g. Dikarya), kinetoplastids, microsporidians, and some other protists. In such genomes intronIC should treat weak motif or snRNA hits as false positives; the adjudicator suppresses their calls only when the species-level evidence reaches NOT_DETECTED.
The minimum number of U2-type introns (= 200) a genome needs before a per-genome EVT tail can be fit; below it the genome is LOW_N → UNASSESSABLE (per-intron P_motif still reported, no bearer/loss call). Opt-in --adjudicator-min-u2 <N> lowers it for a known-representative small set.
The per-species gate the adjudicator assigns over a genome's P_motif calls, one of DETECTED / INCONCLUSIVE / NOT_DETECTED / UNASSESSABLE. It names the motif evidence for a U12-type population, not biological bearer/loss truth, and only NOT_DETECTED suppresses calls (see Adjudicator design).
motif_category value meaning a U12-type population is present by motif — either because z_excess ≥ 5.50 (count gate) or because the genome has a few genuinely strong calls (strength gate). It calls for downstream corroboration, not a proof of function.
motif_category value meaning the adjudicator abstains. Its strong-motif introns still receive type_id = u12; corroborate the species downstream. Two situations produce it:
-
z_excesslies in the[2.60, 5.50)gap and the genome is not strength-rescued — abstention within support; and -
(v3.1) the genome has only one or two strong calls, too few for the strength gate to run, but its strongest call is one its own U2-type background cannot account for (
min(bg_fdr) ≤ 3e-3) — the low-k escape hatch. Before v3.1 these genomes fell through toNOT_DETECTEDand had every score zeroed.
Both mean the same thing operationally — abstain, corroborate downstream — but note that a genome labelled INCONCLUSIVE no longer implies 2.60 < z_excess < 5.50.
motif_category value meaning the genome's strong calls are consistent with its own U2-type background. It is the only category that suppresses calls (all calls zeroed, type_id = u2). It is not itself a loss call — motif-silent divergent bearers can land here, where motif evidence alone is insufficient and orthogonal evidence (e.g. minor-snRNA presence or phylogenetic context) is needed to resolve them.
motif_category value when a per-genome U2-type null cannot be referenced (LOW_N or a degenerate tail). Per-intron P_motif is reported as-is with no species call; calls are not suppressed.
The primary driver of the strength gate: the per-genome Gumbel tail probability that the genome's own U2-type background produces a maximum as strong as its call cs_p95. p_gumbel_p95 ≤ 0.01 — a ~1% one-sided outlier test evaluated at the call upper tail (p95) — rescues few-but-strong divergent bearers to DETECTED. p95 is chosen because bearer recovery saturates there (median 12/46 → p95 36/46) and the worst loss reaches only 0.044.
The count-based significance behind z_excess: given the number of strong calls a genome's U2-type tail would be expected to throw up by chance, how significant is the observed excess. Treating the strong-call count as a Poisson variable referenced to the genome's own tail makes z_excess robust across genome sizes.
The output-level layer that sits on top of P_motif and makes the per-species U12-type population call. It emits two numbers, reported separately: the continuous z_excess (with a strength gate) and the categorical motif_category. All v3 per-species adaptation lives here; the runtime scope is motif-only (no phylogeny or snRNA). See Technical algorithm.
The adjudicator's primary population statistic: the Poisson significance of the count of strong calls (P_motif ≥ 0.9) against what the genome's own U2-type EVT tail predicts. A loss's motif-strong false positives are the U2-type tail (z_excess ≈ 0); a genuine bearer has a separate population the tail cannot explain (large z_excess). Because it is count-based it can miss a bearer with only a few strong calls, which is what the strength gate (p_gumbel_p95) recovers.
The 0–100 calling scale, adjusted_score = 100·q_eff·P_motif (q_eff ∈ {0,1}, 0 only when motif_category = NOT_DETECTED). In any non-NOT_DETECTED genome this equals 100·P_motif; in a NOT_DETECTED genome all calls are zeroed. adjusted_score ≥ 50 ⟺ type_id = u12. This is the adjudicated calling scale, not the removed v2.7 discount penalty. See Output-files.
The number of introns in the intron's parent transcript (the (N) in an intron label). Zero-length annotation artifacts are excluded from this count.
The intron's CDS-relative position: (CDS length upstream of the intron) / (total CDS length), in [0, 1] (e.g. 0.5 for an intron at the midpoint of the coding sequence). Normalizing by the CDS rather than the full mature transcript avoids a 3′UTR-length bias that otherwise pushes introns toward the 5′ end. Exon-only (UTR) introns have no CDS position and are reported as NA.
A call with rel_score > 0, equivalently P_motif > 0.9 in a non-NOT_DETECTED genome — the strong-motif subset of U12-type calls. A u12 type_id alone may include intermediate-confidence calls; filter on rel_score > 0 for the HC subset.
Back-compatibility columns from the superseded q·P_motif design, removed from score_info.iic in v3.1.0 (38 → 35 columns). All four were deterministic functions of P_motif + motif_category: q ∈ {0,1} was a per-species constant (0 only when motif_category = NOT_DETECTED), P_adj = q·P_motif, and P_adj_lo/P_adj_hi were bit-identical to P_adj once the gate became binary. adjusted_score is now computed directly as 100·q_eff·P_motif. Read P_motif + motif_category. Note that type_id moved from column 38 to 35 as a result, so parse by header name rather than position.
Report-only per-intron column (score_info.iic) added in v3.1.0: for each strong call (P_motif ≥ 0.90), the fraction of the calls at least that strong which the genome's own fitted U2-type background tail already accounts for — a Benjamini–Hochberg q-value, monotone in call strength. It is the per-intron decomposition of z_excess, which evaluates the same arithmetic once at the call core. Read ~1 as "indistinguishable from this genome's background" and ≪1 as "the background cannot account for it"; nan for non-calls (those are the background — rank them with P_motif). It is populated even on NOT_DETECTED genomes whose scores are zeroed, which is its main use: it distinguishes a suppressed genome whose calls are uniformly background from one holding a genuine outlier, a distinction neither z_excess nor the saturating P_motif can express. No intron's own call depends on its own bg_fdr; the per-genome minimum drives the low-k escape hatch.
rel_score = adjusted_score − 90 (range ≈ [−90, +10]) — signed and centered on the 90% high-confidence threshold, not on the U12-type/U2-type boundary. rel_score > 0 marks a high-confidence U12-type call; rel_score < 0 means "below the 90% threshold", not "anti-U12-type". See Output-files.
The per-species adjudicator sidecar emitted by the pmotif_adjudicated bundle, holding the fitted U2-type tail model and the numbers behind motif_category (z_excess, cs_p95, p_gumbel_p95, the per-call margins, and the EVT fit parameters). It is written from the same margins the adjudicator scored, so it is bit-exact to the gate and drives the .plot.tail_model.iic.png diagnostic.
The authoritative per-intron classification, u12 or u2 (never NA). type_id = u12 iff P_motif ≥ 0.5 and motif_category ≠ NOT_DETECTED. It appears in score_info.iic and meta.iic.
A threshold-free measure of ranking quality (1.0 = perfect, 0.5 = chance). The headline v3 result is leave-clade-out AUC on loss-species false-positive discrimination: 0.916 for raw motif features versus 0.786 for z-normalized features. This is the primary reason z-normalization was removed, since it manufactures loss-species false positives.
The regularized blend used in background correction: an empirical species composition is mixed with the human U2-type prior by weight w = n/(n + n0), where n is the species' intron count and n0 (default 1000) sets how much data is needed before the empirical estimate dominates (at n = n0, w = 0.5). It keeps low-count species from over-fitting a noisy composition.
A clade is a monophyletic group of species; the training corpus spans 14 of them. Leave-clade-out validation holds an entire clade out of training and evaluates on it, testing genuine cross-lineage generalization (rather than leaking near-relatives between train and test). intronIC reports leave-clade-out AUC, ECE, and adjudicator cross-error rates.
Assigning training-set U12-type/U2-type labels from cross-species evidence — aligning orthologous genes across species and using conserved intron positions and motifs to assign a confident U12-type or U2-type label — rather than from any single genome (Larue & Roy 2023). It is the label source for the v3 corpus (inclusion required ≥3 species per ortholog group across ≥2 phyla) and yields labels robust to single-species motif noise.
The average gap between predicted probability and observed frequency, measuring how well-calibrated P_motif is. The frozen Platt fit achieves leave-clade-out ECE ≈ 7×10⁻⁴, i.e. a stated probability of 0.9 corresponds closely to a 90% empirical U12-type rate.
A corpus cleaning step that drops U12-type records supported in only a single lineage and showing motif decay, likely lineage-specific U12→U2 conversion events rather than conserved minor introns (37 records dropped in the May 2026 build). It removes probable label noise before training.
The v3 training pool the bundled model was fit on: 41,333 introns = 10,003 U12-type + 31,330 U2-type, drawn from 97 species across 14 clades (post-singleton-decay filter), with labels assigned by comparative genomics labeling. A further 5 species (790 U12-type positives) are held out for evaluation, on which the model reaches F1 = 1.000 and Brier ≈ 4×10⁻⁶. See Training-data-and-PWMs.
- Alioto TS. (2007) U12DB: a database of orthologous U12-type spliceosomal introns. Nucleic Acids Research 35(Database):D110–D115. doi:10.1093/nar/gkl796
- Coles S. (2001) An Introduction to Statistical Modeling of Extreme Values. Springer Series in Statistics, Springer. doi:10.1007/978-1-4471-3675-0
- Larue GE, Roy SW. (2023) Where the minor things are: a pan-eukaryotic survey suggests neutral processes may explain much of minor intron evolution. Nucleic Acids Research 51(20):10884–10908. doi:10.1093/nar/gkad797
- Moyer DC, Larue GE, Hershberger CE, Roy SW, Padgett RA. (2020) Comprehensive database and evolutionary dynamics of U12-type introns. Nucleic Acids Research 48(13):7066–7078. doi:10.1093/nar/gkaa464
- Platt JC. (1999) Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. In Advances in Large Margin Classifiers, MIT Press, pp. 61–74.
- Zeng Y, Fair BJ, Zeng H, Krishnamohan A, Hou Y, Hall JM, Ruthenburg AJ, Li YI, Staley JP. (2022) Profiling lariat intermediates reveals genetic determinants of early and late co-transcriptional splicing. Molecular Cell 82(24):4681–4699.e8. doi:10.1016/j.molcel.2022.11.004