Skip to main content

​
Overview

The predict module generates TRIFID scores for transcript isoforms using a trained model. It handles multiple genome assemblies and releases, with automatic feature preprocessing and missing value handling.

​
Command-Line Usage

python -m trifid.models.predict \
    --config config/config.yaml \
    --features config/features.yaml \
    --model models/trifid.v_1_0_0.pkl \
    --assembly GRCh38 \
    --release g27

​
Command-Line Arguments

​
-a, --assembly
string
default:"GRCh38"
Genome assembly version. Supported assemblies:
  • Human: GRCh38, GRCh37
  • Mouse: GRCm39, GRCm38
  • Rat: Rnor_6.0
  • Zebrafish: GRCz11
  • Pig: Sscrofa11.1
  • Chimpanzee: Pan_tro_3.0
  • Chicken: GRCg6a
  • Cow: ARS-UCD1.2
  • Fly: BDGP6
  • Worm: WBcel235
​
-c, --config
string
default:"config/config.yaml"
Path to configuration file containing genome and annotation settings
​
-f, --features
string
default:"config/features.yaml"
Path to features selected description YAML file
​
-m, --model
string
default:"models/trifid.v_1_0_4.pkl"
Path to pretrained model pickle file
​
-r, --release
string
default:"g27"
Genome release version (e.g., g27, r110, e104)

​
Core Function

​
make_predictions()

Generates TRIFID predictions for transcript isoforms. 
from trifid.models.predict import make_predictions

make_predictions(
    features=feature_list,
    ids=identifier_list,
    config=config_dict,
    model=trained_model,
    assembly="GRCh38",
    release="g27"
)

​
Parameters

​
features
list
required
List of feature column names to use for prediction
​
ids
list
required
List of identifier column names (e.g., gene_id, transcript_id)
​
config
dict
required
Configuration dictionary containing:
  • annotation.genome_version: Genome assembly version
  • annotation.db: Database directory name
​
model
object
required
Trained scikit-learn model object loaded from pickle file
​
assembly
string
required
Genome assembly version identifier
​
release
string
required
Genome release identifier

​
Returns

Writes predictions to trifid_predictions.tsv.gz with the following columns:
​
gene_id
string
Ensembl gene identifier
​
gene_name
string
Gene symbol/name
​
transcript_id
string
Ensembl transcript identifier
​
translation_id
string
Ensembl protein identifier
​
flags
string
Annotation flags
​
ccdsid
string
CCDS identifier (if available)
​
appris
string
APPRIS annotation
​
ann_type
string
Annotation type
​
length
integer
Transcript length
​
trifid_score
float
Raw TRIFID score (probability from 0-1)
​
norm_trifid_score
float
Normalized TRIFID score within gene

​
Assembly-Specific Preprocessing

The function automatically handles missing values and feature differences across assemblies:

​
Human Assemblies

GRCh38 (Gencode):
  • No special preprocessing required
  • All features available
GRCh37 (Gencode):
  • Fills missing values with -1
GRCh38/GRCh37 (RefSeq):
  • Fills RefSeq-unavailable features with -1
  • Handles CCDS binary encoding
  • Fills corsair features with 0
  • Fills RNA2sj features with -1 (GRCh37 only)

​
Model Organism Assemblies

Mouse (GRCm39, GRCm38):
  • Fills RNA2sj features with -1
Rat (Rnor_6.0), Zebrafish (GRCz11), Pig (Sscrofa11.1), Chimpanzee (Pan_tro_3.0):
  • Removes CCDS feature
  • Loads ACDS (Alternative CCDS) data
  • Fills RNA2sj features with -1
Chicken (GRCg6a), Cow (ARS-UCD1.2), Fly (BDGP6):
  • Fills all missing values with -1
Worm (WBcel235):
  • Fills all missing values with -1 or 0

​
Example Workflows

​
Basic Prediction

import pickle
import pandas as pd
from trifid.models.predict import make_predictions
from trifid.utils.utils import parse_yaml

# Load configuration
config = parse_yaml("config/config.yaml")
df_features = pd.DataFrame(parse_yaml("config/features.yaml"))

# Load model
model = pickle.load(open("models/trifid.v_1_0_4.pkl", "rb"))

# Extract feature and identifier lists
features = df_features[
    ~df_features["category"].str.contains("Identifier")
]["feature"].values
ids = df_features[
    df_features["category"].str.contains("Identifier")
]["feature"].values

# Generate predictions
make_predictions(
    features=features,
    ids=ids,
    config=config,
    model=model,
    assembly="GRCh38",
    release="g27"
)

​
Predict for Multiple Assemblies

# Human GRCh38
python -m trifid.models.predict \
    --assembly GRCh38 \
    --release g27

# Human GRCh37
python -m trifid.models.predict \
    --assembly GRCh37 \
    --release g27

# Mouse GRCm39
python -m trifid.models.predict \
    --assembly GRCm39 \
    --release m27

​
Custom Model Path

python -m trifid.models.predict \
    --config config/config.yaml \
    --features config/features.yaml \
    --model models/custom_model.pkl \
    --assembly GRCh38 \
    --release g27

​
Input Requirements

​
TRIFID Database

The input database (trifid_db.tsv.gz) must contain:
  1. Identifier columns: gene_id, gene_name, transcript_id, translation_id, ccdsid
  2. Feature columns: All features specified in features YAML
  3. Annotation columns: flags, appris, ann_type

​
Configuration File

The config YAML must specify: 
annotation:
  genome_version: "GRCh38"
  db: "g27"

​
Features File

The features YAML must include: 
- feature: "gene_id"
  category: "Identifier"
- feature: "transcript_length"
  category: "Structural"
  refseq: "y"  # or "n" for RefSeq-incompatible features

​
Output Format

Predictions are saved to: 
data/genomes/{assembly}/{release}/trifid_predictions.tsv.gz
The output file is a compressed TSV with headers and contains one row per transcript isoform.

​
Performance Considerations

  • The function processes the entire TRIFID database for the specified assembly
  • Memory usage depends on genome size (human: ~200K transcripts)
  • Prediction time: ~1-10 seconds for human genome with RandomForest model
  • Output file size: ~5-15 MB compressed
  • train: Model training
  • interpret: Model interpretation and feature importance
  • select: Model selection and evaluation

Build docs developers (and LLMs) love

Get started for freeTalk to us