TIScalling - TIS Prediction Using Machine Learning

TIScalling is an open-source project dedicated to predicting non-canonical Translation Initiation Sites (TIS) in transcriptomic data across various species. The project uses advanced machine learning techniques to develop a robust classifier that accurately identifies non-canonical TIS locations based on sequence data and other relevant features.

Table of Contents

1. Introduction
2. Features and Dataset
3. Installation
4. Usage
5. Model Training
6. Evaluation
7. License

Introduction

Translation Initiation Sites (TIS) are critical for protein synthesis, and accurately identifying non-canonical sites can advance our understanding of gene expression regulation across different species. This project leverages machine learning to predict non-canonical TIS, emphasizing scalability and accuracy.

Features and Dataset

Features

• Kozake: Features based on the Kozak sequence context for translation initiation.
• Noderer: Features derived from Noderer score metrics.
• PWM: Position Weight Matrix features based on TIS and their neighboring sequences.
• TIS codon usage: Codon usage statistics specific to TIS.
• CDS: Coding sequence-specific features.
• RNA: RNA secondary structure and accessibility features.
• Kmer: k-mer frequency features.

Dataset

This project uses structured datasets to train and evaluate the TIS prediction models. The dataset consists of:

• Inputs for Model Training:

  • Labeled examples of TIS (both positive and negative) in TSV format.
  • Transcript sequences in TSV format, providing the context for TIS identification.

• Inputs for TIS Prediction:

  • DNA sequences in FASTA format for identifying TIS in new data.

Example Files

The following example files are automatically generated in the input_data/ folder after installation: *

• Tomato_TP_and_TN_data.txt: Contains true positive and true negative TIS examples.
• Tomato_seq.txt: Includes transcript sequences used for feature extraction.
• predict_input.fa: FASTA file used as input for TIS prediction.

These files serve as a starting point for processing and feature generation during model training and prediction. - Tomato_TP_and_TN_data.txt: True positive and true negative TIS sites.

• Tomato_seq.txt: Transcript sequences.
• predict_input.fa: Input data for prediction in FASTA format.

Installation

# Clone the repository
git clone https://github.com/yenmr/TIScalling.git
cd TIScalling

# Create Conda environment
conda create -n TIScalling python=3.10
conda activate TIScalling

# Install dependencies
pip install -r requirements.txt

Usage

Preprocess Data

Prepare PWM and Codon Usage Table for generating features

# Prepare PWM 
python script/PWM_generator.py -seq input_data/Tomato_seq.txt -TP_TN_data input_data/Tomato_TP_and_TN_data.txt -filename PWM_Sl_5UTRATG.txt -TP_type 2
python script/PWM_generator.py -seq input_data/Tomato_seq.txt -TP_TN_data input_data/Tomato_TP_and_TN_data.txt -filename PWM_Sl_CDSATG.txt -TP_type 3
python script/PWM_generator.py -seq input_data/Tomato_seq.txt -TP_TN_data input_data/Tomato_TP_and_TN_data.txt -filename PWM_Sl_5UTRnonATG.txt -TP_type 4
python script/PWM_generator.py -seq input_data/Tomato_seq.txt -TP_TN_data input_data/Tomato_TP_and_TN_data.txt -filename PWM_Sl_CDSnonATG.txt -TP_type 5

# -TP_type specifies the TIS category:
# 2: 5UTR_ATG, 3: CDS_ATG, 4: 5UTR_nonATG, 5: CDS_nonATG

# Output Directory: The generated PWM files will be stored in the `PWM_library` folder.

# Prepare Codon Usage Table
python script/coden_usage_generator.py -TP_TN_data input_data/Tomato_TP_and_TN_data.txt -seq input_data/Tomato_seq.txt -TP_type 2 -filename TIS_codon_Sl_5UTRATG.txt -sp Sl
python script/coden_usage_generator.py -TP_TN_data input_data/Tomato_TP_and_TN_data.txt -seq input_data/Tomato_seq.txt -TP_type 3 -filename TIS_codon_Sl_CDSATG.txt -sp Sl
python script/coden_usage_generator.py -TP_TN_data input_data/Tomato_TP_and_TN_data.txt -seq input_data/Tomato_seq.txt -TP_type 4 -filename TIS_codon_Sl_5UTRnonATG.txt -sp Sl
python script/coden_usage_generator.py -TP_TN_data input_data/Tomato_TP_and_TN_data.txt -seq input_data/Tomato_seq.txt -TP_type 5 -filename TIS_codon_Sl_CDSnonATG.txt -sp Sl

# -TP_type specifies the TIS category:
# 2: 5UTR_ATG, 3: CDS_ATG, 4: 5UTR_nonATG, 5: CDS_nonATG

# Output Directory: The generated codon usage files will be stored in the `TIS_codon_library/` folder.

Generate Features

python script/get_features.py -TP_TN_data input_data/Tomato_TP_and_TN_data.txt -seq input_data/Tomato_seq.txt -TIS_sp Sl -TIS_code 2
python script/get_features.py -TP_TN_data input_data/Tomato_TP_and_TN_data.txt -seq input_data/Tomato_seq.txt -TIS_sp Sl -TIS_code 3
python script/get_features.py -TP_TN_data input_data/Tomato_TP_and_TN_data.txt -seq input_data/Tomato_seq.txt -TIS_sp Sl -TIS_code 4
python script/get_features.py -TP_TN_data input_data/Tomato_TP_and_TN_data.txt -seq input_data/Tomato_seq.txt -TIS_sp Sl -TIS_code 5

# -TIS_code specifies the TIS category:
# 2: 5UTR_ATG, 3: CDS_ATG, 4: 5UTR_nonATG, 5: CDS_nonATG

# Output Directory: The generated feature files will be stored in the `feature/` folder.

Train Model

python script/train_model.py -feature Features_Sl_CDSATG.txt
python script/train_model.py -feature Features_Sl_5UTRATG.txt
python script/train_model.py -feature Features_Sl_CDSnonATG.txt
python script/train_model.py -feature Features_Sl_5UTRnonATG.txt

# Output Directory: Trained models and associated files are stored in subdirectories named `Model_[feature name]` within the `model/` folder. The outputs include:
#
# - `Feature_scaler.pkl` and `Feature_scaler_attributes.tsv`: Feature scaling factors in PKL and TSV formats, respectively.
# - `Important_features.tsv`: Contains the importance of each feature.
# - `[classifier]_model.pkl`: Model file for the specified classifier (e.g., GradientBoosting, LogisticRegression, RandomForest, SVM) in PKL format.
# - `[classifier]_model_importances.tsv`: Importance of each feature for the corresponding classifier.
# - `Model_evaluation_results.tsv`: Evaluation results of the trained models.

Cross Model Evaluation

python script/test_model.py -feature Features_Sl_5UTRATG.txt -model Model_Sl_5UTRATG,Model_Sl_CDSATG,Model_Sl_5UTRnonATG,Model_Sl_CDSnonATG

# Output Files: The evaluation results will be saved as `Cross_evaluation_result_[input feature name].tsv`.

TIS Prediction

python script/predict_TIS.py -fasta input_data/predict_input.fa -models Model_Sl_5UTRATG,Model_Sl_CDSATG,Model_Sl_5UTRnonATG,Model_Sl_CDSnonATG

# Output Files: The prediction output for each FASTA sequence will be saved as `Predict_[FASTA header].tsv`.
# Each FASTA sequence generates one TSV file.

Model Training

The model is trained using the following steps:

1. Feature extraction.
2. Train-test split.
3. Hyperparameter optimization using grid search or Bayesian optimization.
4. Model evaluation on a separate validation set.

Supported machine learning algorithms include:

• Random Forest (RF)
• Gradient Boosted Trees (GB)
• Support Vector Machines (SVM)
• Logistic Regression (LR)

Evaluation

Evaluation metrics include:

• F1: Harmonic mean of precision and recall.
• AUCROC: Area under the Receiver Operating Characteristic curve.
• MCC: Matthews Correlation Coefficient for balanced evaluation of predictions.
• Accuracy: Overall prediction correctness.
• Recall: Proportion of actual TIS correctly identified.
• Precision: Proportion of true positive TIS predictions.
• ConfusionMatrix: Detailed classification performance representation.

Results are visualized using confusion matrices, ROC curves, and feature importance plots.

License

This project is licensed under the MIT License. See the LICENSE file for details.