Skip to main content

​
Overview

TRIFID provides pre-trained machine learning models that can predict the functional relevance of splice isoforms. These models have been trained on large-scale proteomics data and validated across multiple species.

​
Available Models

​
Human Model (v1)

Training Data: GENCODE Release 27 (GRCh38.p10) Features: 45 predictive features Release Date: March 10, 2021 Download: TRIFID v1 Model (pickle format) Training Set: GENCODE 27 training data Species Applicability: Homo sapiens (human)

​
Enhanced Models (v2)

Training Data: Extended proteomics evidence with additional features Features: 47 predictive features (2 additional features) Release Date: September 2022 Species Applicability:
  • Homo sapiens (Human-specific model)
  • Mus musculus (Mouse-specific model)
  • Other vertebrates (Vertebrates model)
  • Invertebrates (Invertebrates model)
The v2 models include two additional features that improve prediction accuracy, particularly for minor isoforms.

​
Model Architecture

TRIFID uses a gradient boosting machine learning approach that combines:
  • Accuracy: High-confidence predictions validated against proteomics data
  • Interpretability: SHAP values for feature importance and local predictions
  • Reproducibility: Complete training pipeline and configuration files available

​
Model Input Features

The models use 47 predictive features across multiple categories:
  1. Structural features: Protein domain integrity (Pfam), sequence length
  2. Conservation features: PhyloCSF scores, ALT-Corsair evolutionary age
  3. Expression features: Splice junction coverage (QSplice), tissue specificity
  4. Annotation features: APPRIS principal isoform scores, CDS completeness
For a complete feature list, see the feature documentation.

​
Model Output

TRIFID Score: Probability (0-1) representing functional relevance
  • 0.0-0.3: Low functional probability (likely neutral evolution)
  • 0.3-0.7: Uncertain functional relevance
  • 0.7-1.0: High functional probability (likely under purifying selection)

​
Using the Model

​
Loading a Pre-trained Model

import pickle
import pandas as pd

# Load the model
with open('trifid_model.pkl', 'rb') as f:
    model = pickle.load(f)

# Load predictions
predictions = pd.read_csv(
    'data/genomes/GRCh38/g27/trifid_predictions.tsv.gz',
    compression='gzip',
    sep='\t'
)

​
Example: Querying FGFR1 Isoforms

# Select a gene to explore
gene_name = 'FGFR1'

# Filter predictions
fgfr1_isoforms = predictions.loc[
    predictions['gene_name'] == gene_name,
    ['transcript_id', 'gene_name', 'trifid_score', 'appris', 'sequence']
]

print(fgfr1_isoforms)
Output:
Transcript IDGene NameAPPRIS LabelLengthTRIFID Score
ENST00000447712FGFR1PRINCIPAL:38220.87
ENST00000356207FGFR1MINOR7330.60
ENST00000397103FGFR1MINOR7330.01
ENST00000619564FGFR1MINOR2280.00

​
Interpreting Predictions with SHAP

TRIFID includes SHAP (SHapley Additive exPlanations) values for model interpretability: 
from trifid.models.interpret import explain_prediction

# Load SHAP values
df_shap = pd.read_csv('shap_values.tsv.gz', compression='gzip', sep='\t')

# Explain a specific isoform
explain_prediction(df_shap, model, features, 'ENST00000356207')
This generates a waterfall plot showing which features contribute most to the prediction for that specific isoform.

​
Model Training Pipeline

If you want to train TRIFID on custom data or reproduce the training process:

​
1. Prepare the Dataset

python -m trifid.data.make_dataset
This creates a complete dataset with all 47 predictive features.

​
2. Train the Model

python -m trifid.model.train
The training script:
  • Loads the training set (proteomics-validated isoforms)
  • Performs hyperparameter optimization
  • Trains the gradient boosting model
  • Saves the model in pickle format

​
3. Generate Predictions

python -m trifid.model.predict
Applies the trained model to all isoforms in the genome annotation.

​
Model Configuration

The training pipeline is controlled by configuration files in the config/ directory:
  • config.yaml: File paths and pipeline parameters
  • features.yaml: Feature definitions, categories, and species support

​
Model Performance

The TRIFID model has been validated to show:
  • High concordance with proteomics detection
  • Predicted functional isoforms show measurable cross-species conservation
  • Exons from high-scoring isoforms are under purifying selection
  • Low-scoring isoforms show evidence of neutral evolution
For detailed performance metrics, see the original publication.

​
Tutorials and Examples

Comprehensive tutorials are available as Jupyter notebooks:

​
Model Applicability

While TRIFID was developed for the human genome, the models can be applied to:
Use the Human-specific model (v1 or v2) for GRCh37 or GRCh38 assemblies with GENCODE or RefSeq annotations.
Use the Mouse-specific model (v2) for GRCm38 or GRCm39 assemblies with GENCODE annotations.
Use the Vertebrates model (v2) for rat, zebrafish, chicken, chimpanzee, pig, cow, and macaque.
Use the Invertebrates model (v2) for fruitfly and worm. Note that prediction accuracy may be lower due to evolutionary distance.
For species not listed in the available predictions, model applicability depends on the availability of required data sources (APPRIS, PhyloCSF, etc.). Contact the developers for guidance.

​
Citation

If you use TRIFID models in your research, please cite: 
@article{10.1093/nargab/lqab044,
    author = {Pozo, Fernando and Martinez-Gomez, Laura and Walsh, Thomas A and 
              Rodriguez, José Manuel and Di Domenico, Tomas and Abascal, Federico and 
              Vazquez, Jesús and Tress, Michael L},
    title = "{Assessing the functional relevance of splice isoforms}",
    journal = {NAR Genomics and Bioinformatics},
    volume = {3},
    number = {2},
    year = {2021},
    doi = {10.1093/nargab/lqab044}
}

Build docs developers (and LLMs) love

Get started for freeTalk to us