Skip to main content

​
Feature Categories

TRIFID integrates 45+ features spanning multiple biological dimensions to predict isoform functionality. These features are organized into seven main categories:

Annotation

12 features from transcript metadata

Evolution

8 conservation and evolutionary scores

Structure

14 protein structure and domain features

Splicing

6 alternative splicing impact features

Expression

2 RNA-seq junction support features

Derived

Multiple normalized and delta scores
All features are extracted from the unified TRIFID database (trifid_db.tsv.gz) which integrates data from APPRIS, PhyloCSF, Pfam, RNA-seq, and genome annotations.

​
Annotation Features

Metadata and annotation quality indicators from GENCODE/Ensembl.

​
basic

  • Description: Simplified, high-quality subset of GENCODE annotations
  • Type: Binary (0/1)
  • Source: GENCODE Basic tag
  • Interpretation: 1 = transcript in the basic gene annotation set
  • Species Support: GRCh38, GRCm38

​
CCDS

  • Description: Consensus CDS protein set membership
  • Type: Binary (0/1)
  • Source: NCBI CCDS database
  • Interpretation: 1 = transcript has a CCDS identifier (high confidence)
  • Species Support: Cross-species (GRCh38, GRCm38, Rnor_6.0)
  • Code: trifid/data/feature_engineering.py:102-109
The Consensus Coding Sequence (CCDS) project identifies protein-coding regions that are consistently annotated across multiple genome annotation sources. CCDS transcripts are highly confident and experimentally supported.

​
Transcript Support Level (tsl_1 through tsl_6)

  • Description: Confidence in transcript structure based on RNA-seq evidence
  • Type: One-hot encoded (6 binary features)
  • Categories:
    • TSL 1: All splice junctions supported by non-suspect mRNA
    • TSL 2: Best supporting mRNA is suspect OR support from multiple ESTs
    • TSL 3: Only support from single EST
    • TSL 4: Best supporting EST is suspect
    • TSL 5: No mRNA support for structure
    • TSL 6: Not analyzed for support
  • Source: Ensembl transcript quality tags
  • Species Support: GRCh38, GRCm38
  • Code: trifid/data/feature_engineering.py:50-73

​
Annotation Level (level_1, level_2, level_3)

  • Description: Annotation confidence based on evidence type
  • Type: One-hot encoded (3 binary features)
  • Categories:
    • Level 1: Verified loci (highest confidence)
    • Level 2: Manual annotation (Havana/Ensembl)
    • Level 3: Automated annotation
  • Source: GENCODE annotation level
  • Species Support: GRCh38, GRCm38

​
nonsense_mediated_decay

  • Description: Transcript likely targeted for NMD degradation
  • Type: Binary (0/1)
  • Source: Ensembl biotype annotation
  • Interpretation: 1 = contains premature stop codon, likely degraded
  • Species Support: GRCh38, GRCm38, Rnor_6.0, GRCz11

​
StartEnd_NF

  • Description: Start or end codon not found
  • Type: Binary (0/1)
  • Source: GENCODE tag
  • Interpretation: 1 = 5’ or 3’ end incomplete
  • Species Support: GRCh38

​
length

  • Description: Protein length in amino acids
  • Type: Continuous (integer)
  • Source: Protein FASTA sequences
  • Range: Typically 10-5,000+ residues
  • Species Support: Cross-species

​
length_delta_score

  • Description: Length similarity to principal/longest isoform
  • Type: Continuous (0-1)
  • Formula: 1 - abs((ref_length - length) / ref_length)
  • Interpretation: 1.0 = same length as reference, 0.0 = maximally different
  • Code: trifid/utils/utils.py:127-170
length_delta_score captures the biological principle that functional isoforms often have similar lengths to the principal isoform, while short truncations may be non-functional.

​
Evolution Features

Conservation and evolutionary constraint signals.

​
CORSAIR (corsair)

  • Description: Cross-species conservation score
  • Type: Continuous (0-4)
  • Method: Number of vertebrate species with full-length orthologous isoform
  • Source: APPRIS database
  • Interpretation: Higher = more conserved across vertebrates
  • Species Support: Cross-species

​
norm_corsair

  • Description: Gene-normalized CORSAIR score
  • Type: Continuous (0-1)
  • Formula: (corsair - min_gene) / (max(4, max_gene) - min_gene)
  • Interpretation: 1.0 = highest conservation within gene
  • Code: trifid/data/feature_engineering.py:35-47

​
Alt-CORSAIR (corsair_alt)

  • Description: Oldest species with conserved isoform
  • Type: Continuous (0-0.25)
  • Scale: Bilateria (oldest) to recent vertebrates
  • Source: APPRIS database
  • Species Support: Cross-species

​
norm_corsair_alt

  • Description: Gene-normalized Alt-CORSAIR
  • Type: Continuous (0-1)
  • Code: trifid/data/feature_engineering.py:76-92

​
PhyloCSF Features

PhyloCSF measures evolutionary signatures of protein-coding regions using comparative genomics.

​
ScorePerCodon

  • Description: PhyloCSF score normalized per codon
  • Type: Continuous (can be negative)
  • Method: Log-likelihood ratio of coding vs. non-coding evolutionary model
  • Interpretation: Higher = stronger coding signature
  • Species Support: GRCh38, GRCm38

​
norm_ScorePerCodon

  • Description: Gene-normalized ScorePerCodon
  • Type: Continuous (0-1)
  • Handling: Missing values imputed at 3rd percentile
  • Code: trifid/data/feature_engineering.py:94-114

​
PhyloCSF_Psi

  • Description: Length-adjusted PhyloCSF score
  • Type: Continuous
  • Method: Adjusts for coding sequence length
  • Species Support: GRCh38, GRCm38

​
norm_PhyloCSF_Psi

  • Description: Gene-normalized PhyloCSF_Psi
  • Type: Continuous (0-1)

​
RelBranchLength

  • Description: Minimum relative branch length in phylogenetic alignment
  • Type: Continuous (0-1)
  • Method: Lowest exon-based score showing species coverage
  • Interpretation: Higher = better alignment quality
  • Species Support: GRCh38, GRCm38

​
norm_RelBranchLength

  • Description: Gene-normalized RelBranchLength
  • Type: Continuous (0-1)
PhyloCSF features are only available for human (GRCh38) and mouse (GRCm38) assemblies. For other species, these features are imputed with -1.

​
Structure Features

Protein structure predictions from APPRIS methods.

​
firestar

  • Description: Number of functional residues detected
  • Type: Continuous (integer)
  • Method: Identifies ligand-binding and catalytic residues
  • Source: APPRIS firestar module
  • Species Support: Cross-species

​
norm_firestar

  • Description: Gene-normalized firestar score
  • Type: Continuous (0-1)
  • Code: trifid/data/feature_engineering.py:35-47

​
matador3d

  • Description: Exon-structure mapping score
  • Type: Continuous
  • Method: Number of exons mapping to known 3D structures
  • Source: APPRIS Matador3D module
  • Species Support: Cross-species

​
norm_matador3d

  • Description: Gene-normalized Matador3D score
  • Type: Continuous (0-1)

​
SPADE (spade)

  • Description: Pfam domain integrity score
  • Type: Continuous
  • Method: Sum of Pfam alignment bitscores
  • Source: APPRIS SPADE module
  • Interpretation: Higher = more complete domains
  • Species Support: Cross-species

​
norm_spade

  • Description: Gene-normalized SPADE score
  • Type: Continuous (0-1)

​
spade_loss

  • Description: Domain loss relative to principal isoform
  • Type: Continuous (capped at 50)
  • Formula: max_gene_spade - spade
  • Code: trifid/data/feature_engineering.py:130

​
norm_spade_loss

  • Description: Gene-normalized SPADE loss
  • Type: Continuous (0-1)
  • Formula: 1 - norm_spade_loss (inverted)
  • Code: trifid/data/feature_engineering.py:132-133

​
CRASH Scores (crash_p, crash_m)

  • Description: Signal peptide reliability scores
  • crash_p: Secretory signal sequence
  • crash_m: Mitochondrial signal sequence
  • Type: Continuous
  • Source: APPRIS CRASH module
  • Species Support: Cross-species

​
norm_crash_p, norm_crash_m

  • Description: Gene-normalized CRASH scores
  • Type: Continuous (0-1)

​
THUMP (thump)

  • Description: Number of transmembrane helices
  • Type: Integer (0-10+)
  • Source: APPRIS THUMP module
  • Species Support: Cross-species

​
norm_thump

  • Description: Gene-normalized THUMP score
  • Type: Continuous (0-1)

​
Splicing Features

Pfam domain impact from alternative splicing (QPfam analysis).

​
pfam_score

  • Description: Percentage of residues in intact Pfam domains
  • Type: Continuous (0-100)
  • Source: QPfam module
  • Method: Custom Pfam domain analysis pipeline
  • Species Support: Cross-species

​
pfam_domains_impact_score

  • Description: Percentage of intact Pfam domains after splicing
  • Type: Continuous (0-100)
  • Source: QPfam module

​
perc_Damaged_State

  • Description: Percentage of Pfam domains damaged by splicing
  • Type: Continuous (0-100)
  • Method: Domains partially lost or truncated
  • Source: QPfam module

​
perc_Lost_State

  • Description: Percentage of Pfam domains completely lost
  • Type: Continuous (0-100)
  • Source: QPfam module

​
norm_Lost_residues_pfam

  • Description: Normalized count of lost Pfam residues
  • Type: Continuous (0-1)
  • Special handling: Values < 10 residues set to 0
  • Code: trifid/data/feature_engineering.py:128

​
norm_Gain_residues_pfam

  • Description: Normalized count of gained Pfam residues
  • Type: Continuous (0-1)
  • Source: QPfam module
QPfam is a custom module that analyzes how alternative splicing affects Pfam protein domains. It identifies:
  • Intact domains: Fully preserved
  • Damaged domains: Partially truncated
  • Lost domains: Completely removed
  • Gained domains: New domains added
This provides crucial information about functional consequences of splicing.

​
Expression Features

RNA-seq splice junction support (QSplice analysis).

​
RNA2sj_cds

  • Description: Splice junction read support ratio
  • Type: Continuous (0-1+)
  • Method: Min junction reads / average CDS junction reads
  • Source: QSplice module analyzing GTEx RNA-seq data
  • Interpretation: Higher = better junction support
  • Species Support: GRCh38 only
  • Code: trifid/data/feature_engineering.py:137-155

​
norm_RNA2sj_cds

  • Description: Gene-normalized RNA2sj_cds
  • Type: Continuous (0-1)
RNA-seq features (RNA2sj_cds) are only available for human GRCh38. For other assemblies, these are imputed with -1.
Low RNA2sj_cds scores may indicate:
  • Poorly supported splice junctions
  • Annotation artifacts
  • Tissue-specific or rare isoforms
  • Potential non-functional transcripts

​
Feature Engineering

TRIFID applies sophisticated transformations to raw features.

​
Group Normalization

# From trifid/utils/utils.py:359-397
def group_normalization(df, features, nmax=0, nmin=0, groupby_feature='gene_id'):
    # Min-max normalization within each gene
    df[f'norm_{feature}'] = df.groupby(groupby_feature)[feature].transform(
        lambda x: (x - min(nmin, x.min())) / (max(nmax, x.max()) - min(nmin, x.min()))
    )
Rationale: Captures relative differences between isoforms of the same gene rather than absolute scores. Example: A CORSAIR score of 3.5 may be high for one gene but low for another. Normalization reveals which isoform is most conserved within its gene context.

​
Delta Scoring

# From trifid/utils/utils.py:127-170
def delta_score(df, features, mode='appris'):
    # Calculate difference from reference (APPRIS principal)
    df[f'{feature}_reference'] = df.groupby(['gene_id'])[feature].transform('first')
    df[f'{feature}_delta_score'] = 1 - abs(
        (df[f'{feature}_reference'] - df[feature]) / df[f'{feature}_reference']
    )
Purpose: Measures similarity to the principal isoform (typically the most functional).

​
Fragment Correction

# From trifid/utils/utils.py:173-223
def fragments_correction(df_iso, features):
    # Adjusts scores for incomplete transcript fragments
    # by referencing complete homologous transcripts
Purpose: Prevents fragment isoforms from receiving artificially low scores due to incompleteness.

​
One-Hot Encoding

Categorical features (TSL, Level) are one-hot encoded: 
# From trifid/utils/utils.py:480-498
def one_hot_encoding(df, features):
    # Converts categories to binary indicator variables
Example: TSL value of “2” becomes: 
tsl_1=0, tsl_2=1, tsl_3=0, tsl_4=0, tsl_5=0, tsl_6=0

​
Missing Data Handling

Different strategies for different assembly types (see trifid/models/predict.py:32-136):
  • GRCh38 Ensembl: Most complete, minimal imputation
  • GRCh37: Fill missing columns with -1
  • RefSeq annotations: CCDS computed from ccdsid, missing features = -1
  • Mouse/Rat/Other species: RNA-seq features = -1, CCDS handled separately
  • Non-mammalian species: More extensive imputation with -1 or 0

Imputation Strategy

  • -1: Feature unavailable for this assembly (not missing data)
  • 0: Biological absence (e.g., no CCDS identifier, no junction support)
  • Percentile: PhyloCSF features imputed at 3rd percentile
  • Gene-level: Fragment correction uses within-gene imputation

​
Feature Importance

Top features contributing to TRIFID predictions (from SHAP analysis):
  1. norm_corsair: Cross-species conservation
  2. length_delta_score: Length similarity to principal
  3. norm_ScorePerCodon: PhyloCSF coding strength
  4. norm_spade: Pfam domain integrity
  5. CCDS: Consensus coding sequence membership
  6. tsl_1: Highest transcript support level
  7. norm_firestar: Functional residues
  8. pfam_score: Domain coverage
  9. norm_RNA2sj_cds: Junction support
  10. perc_Lost_State: Domain loss percentage
See the Interpretability page for detailed SHAP analysis of feature contributions.

​
Feature Summary Table

CategoryCountKey Features
Annotation12CCDS, TSL, Level, basic, length
Evolution8CORSAIR, PhyloCSF (3), RelBranchLength
Structure14SPADE, firestar, Matador3D, CRASH, THUMP
Splicing6Pfam domain impacts, residue gains/losses
Expression2RNA-seq junction support
Derived20+Normalized versions of above
TOTAL45+

​
Next Steps

Interpretability

Learn how SHAP explains which features drive predictions

Feature Configuration

Customize feature selection for your analysis

Data Preparation

Prepare your own feature data

Model Overview

Back to pipeline overview

Build docs developers (and LLMs) love

Get started for freeTalk to us