Skip to content

Technical algorithm

Jump to bottom
Graham Larue edited this page Dec 6, 2025 · 41 revisions

Technical Details

This page provides a detailed technical description of intronIC's algorithm, data flow, and machine learning architecture.

Pipeline Overview

Input: Genome (FASTA) + Annotation (GFF3/GTF) + Species Name

Note: By default, intronIC uses streaming mode, which processes one chromosome at a time for ~85% memory reduction on large genomes.

  1. Stage 1: Intron Extraction

    • Parse annotation hierarchy (gene → transcript → CDS/exon)
    • Infer intron coordinates from exon gaps
    • Extract intron + flanking exon sequences from genome
    • Filter duplicates, short introns, isoforms (configurable)

  2. Stage 2: PWM Scoring

    • Score 5' splice site, branch point, 3' splice site regions
    • Calculate log-odds ratios: log(P(seq|U12) / P(seq|U2))
    • Select best branch point position from search window

  3. Stage 3: Normalization

    • Convert raw log-odds to z-scores via robust scaling
    • Fit on reference data or adapt to experimental data

  4. Stage 4: Classification

    • Feature augmentation (optional composite features)
    • Linear SVM with balanced class weights
    • Probability calibration via Platt scaling
    • Output: P(U12-type) from 0-100%

Output: .meta.iic, .bed.iic, .introns.iic, .score_info.iic, plots

Stage 1: Intron Extraction

Coordinate Inference

Introns are inferred from gaps between consecutive exons (or CDS features) within the same transcript:

Exon 1          Intron 1        Exon 2          Intron 2        Exon 3
[====]----------[      ]--------[====]----------[      ]--------[====]
100-200         201-500         501-700         701-900         901-1100

Priority: CDS features are preferred over exon features when available, as they enable phase calculation.

Filtering Criteria

Filter Default Description
Duplicates Exclude Same coordinates from multiple isoforms
Longest isoform Keep only Can include all with -i
Minimum length 30 bp Adjustable via --min-intron-len
Ambiguous bases Exclude 'N' in scoring regions
Non-canonical Include Exclude with --no-nc

Stage 2: PWM Scoring

Position Weight Matrices

Position weight matrices (PWMs) capture the probability of observing each nucleotide at each position in a motif. intronIC includes PWMs for:

  • U12-type: AT-AC and GT-AG subtypes
  • U2-type: GT-AG and GC-AG subtypes

Each subtype has PWMs for all three regions (5' splice site, branch point, 3' splice site).

Scoring Regions

Default scoring windows relative to splice sites:

Region Start End Length Description
5' SS -3 +9 12 bp Includes last 3 bp of upstream exon
Branch point -55 -5 50 bp Search window from 3'SS
3' SS -6 +4 10 bp Includes first 4 bp of downstream exon

Log-Odds Ratio Calculation

For each region, the raw score is a log-odds ratio:

$$\text{LLR} = \sum_{i=1}^{n} \log\left(\frac{P(b_i | \text{U12 matrix})}{P(b_i | \text{U2 matrix})}\right)$$

Where:

  • $b_i$ is the nucleotide at position $i$
  • Higher positive values favor U12-type
  • Higher negative values favor U2-type
  • Zero means equally likely under both models

Branch Point Selection

The branch point region is searched for the highest-scoring 7-mer matching the U12 branch point consensus (TCCTTAAC or similar). Multiple adenosine position variants are tested, and the highest-scoring position is selected.

Stage 3: Normalization

Why Normalize?

Raw log-odds scores have different ranges and distributions for each region:

  • 5'SS scores might range from -50 to +10
  • BP scores might range from -20 to +5
  • 3'SS scores might range from -5 to +3

Normalization converts these to comparable z-scores for the SVM.

Robust Scaling

intronIC uses robust z-score normalization via sklearn's RobustScaler:

$$z = \frac{\text{raw} - \text{median}}{\text{IQR}}$$

Where IQR is the interquartile range (75th percentile - 25th percentile).

Key properties:

  • Robust to outliers: Median and IQR are resistant to extreme values
  • Centered: Distribution centered around 0 for each feature
  • Comparable scales: All three regions have similar variance after scaling
  • Minimal U12 contamination: Rare U12s (~0.5%) don't significantly affect robust statistics

Normalization Modes

Mode Description Use Case
human Use scales from training (human) data U12-absent species, cross-species
adaptive Refit scales on experimental data Species with different GC content
auto Use human if available in model Default behavior

Note: The pretrained model includes its scaler, so results are reproducible without additional steps when using --normalizer-mode human or auto.

Stage 4: Classification

Feature Space

Base features (3D):

  • five_z_score: 5' splice site z-score
  • bp_z_score: Branch point z-score
  • three_z_score: 3' splice site z-score

Augmented features (optional, 7D default):

  • min_all: min(five_z, bp_z, three_z) — Requires ALL signals strong
  • absdiff_5_bp: |five_z - bp_z| — Penalizes 5'/BP imbalance
  • absdiff_5_3: |five_z - three_z| — Penalizes 5'/3' imbalance
  • absdiff_bp_3: |bp_z - three_z| — Penalizes BP/3' imbalance

The augmented features help the linear SVM reject "one-end-strong" false positives—U2-type introns with one unusually U12-like signal.

Linear SVM

intronIC uses a linear support vector machine (sklearn's LinearSVC):

  • Kernel: Linear (interpretable coefficients)
  • Class weights: Balanced to handle ~0.5% U12 prevalence
  • Regularization: L2 penalty, C optimized via cross-validation
  • Convergence: max_iter=50,000, tol=1e-4

Probability Calibration

The raw SVM outputs decision distances (signed distance from hyperplane). These are converted to probabilities using Platt scaling (sigmoid calibration):

$$P(\text{U12}) = \frac{1}{1 + \exp(A \cdot f(x) + B)}$$

Where $f(x)$ is the decision function output, and $A, B$ are fitted parameters.

Ensemble Training

When --n-models > 1, multiple SVMs are trained with different U2 subsamples:

  1. Each model sees all U12 references but a different 80% of U2 references
  2. Predictions are averaged across models
  3. Reduces variance and improves robustness

Training the Default Model

The default pretrained model was trained on:

Training data:

  • ~400 conserved human U12-type introns (multiple evidence sources)
  • ~20,000 conserved human U2-type introns

Training process:

  1. Score all reference introns with PWMs
  2. Normalize using robust scaling (median/IQR)
  3. Optimize SVM hyperparameters via 5-fold cross-validation
  4. Train final model on all reference data
  5. Calibrate probabilities via cross-validation

Evaluation:

  • Nested cross-validation for unbiased performance estimates
  • Balanced accuracy, F1-score, and PR-AUC reported
  • Validated on held-out human introns and cross-species data

Species-Specific Considerations

GC Content Effects

Species with different GC content than human may show shifted score distributions. Options:

  1. Adaptive normalization: Refit scaler on experimental data (--normalizer-mode adaptive)
  2. Prior adjustment: Adjust base rate expectation (--species-prior)

U12-Absent Lineages

For species known to lack U12-type introns (C. elegans, many fungi):

intronIC -g genome.fa -a annotation.gff -n species \
         --normalizer-mode human --species-prior 1e-6

The reduced prior shifts probability thresholds to minimize false positives.

Cross-Species Performance

The default human-trained model generalizes well to other vertebrates and most eukaryotes with U12-type introns. Performance may degrade for:

  • Very distant lineages (plants, protists)
  • Lineages with unusual U12 motifs
  • Species with extreme GC bias

Consider providing species-specific reference sequences for best results.

Memory and Performance

Standard Mode

Memory scales with annotation density:

  • Loads all intron sequences into memory
  • Human genome (~250k introns): ~12 GB peak

Streaming Mode (--streaming)

Dramatically reduced memory:

  • Writes sequences to temporary SQLite database
  • Keeps only scoring motifs in memory
  • Human genome: ~2-3 GB peak (~85% reduction)
  • Trade-off: Slightly slower I/O

Parallelization

The -p N flag parallelizes PWM scoring:

  • Scoring is CPU-bound and embarrassingly parallel
  • Linear speedup up to ~8-16 cores
  • Diminishing returns beyond that

References

Original intronIC paper:

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

U12-type intron databases:

Alioto TS. (2007) U12DB: a database of orthologous U12-type spliceosomal introns. Nucleic Acids Research 35:D110-D115. doi:10.1093/nar/gkl842

Branch point mapping:

Mercer TR, et al. (2015) Genome-wide discovery of human splicing branchpoints. Genome Research 25:290-303. doi:10.1101/gr.182899.114

SVM probability calibration:

Platt JC. (1999) Probabilistic outputs for support vector machines and comparisons to regularized likelihood methods. Advances in Large Margin Classifiers pp. 61-74.

Clone this wiki locally