TCR/MHC Prediction Models

This repository contains scripts designed to train, evaluate, and visualize models for TCR/MHC interaction predictions. These scripts utilize machine learning classifiers and transform amino acid frequency data to predict TCR/MHC interactions.

Files

• beta.py: This script trains and evaluates three classifiers (Linear SVC, Logistic Regression, and Random Forest) on a dataset of TCR/MHC interactions. It loads the training and testing datasets, performs data transformation, and plots Receiver Operating Characteristic (ROC) curves to evaluate model performance.

• alpha-beta.py: Similar to beta.py, this script includes expanded features in the dataset, working on a larger feature set for training and testing the TCR/MHC interaction prediction models.

Requirements

The following Python packages are required to run the scripts:

• numpy
• pandas
• scikit-learn
• matplotlib

Dataset Format

To use the scripts, the training and testing datasets must be in a specific format, as detailed below. You can use ANARCI (Antibody Numbering and Receptor ClassIfication) to determine the positions in the TCR sequence that correspond to the required columns.

• Columns:
  • Initial columns contain identifiers and gene names:
    • complex.id or MHC Allele Names: Identifiers for each sample.
    • TRA_Gene and TRB_Gene: Gene names for the alpha (TRA) and beta (TRB) chains.
  • TRA and TRB columns: Each subsequent column represents an amino acid at a specific position in the TRA or TRB chain, with columns named as follows:
    • TRA_27, TRA_28, TRA_29, ..., TRA_65 for the alpha chain.
    • TRB_27, TRB_28, TRB_29, ..., TRB_85 for the beta chain.
  • Type: The final column, Type, indicates the binary label for interaction status (0 or 1).

Example

Here is a sample of the dataset format:

complex.id	TRA_Gene	TRB_Gene	TRA_27	TRA_28	...	TRB_84	TRB_85	Type
1	TRAV26-1*01	TRBV13*01	T	I	...	S	D	0
2	TRAV20*01	TRBV13*01	V	S	...	S	D	0

Using ANARCI, you can obtain the specific amino acid positions for your TCR sequences and map them to the correct columns (TRA_27, TRA_28, ..., TRB_85) in the dataset to ensure compatibility with the scripts.

Usage

Both scripts accept command-line arguments for input paths and parameters. The general syntax is:

python script_name.py <train_path> <test_path> [--output_img <filename>]

• train_path: Path to the training dataset in CSV format.
• test_path: Path to the testing dataset in CSV format.
• --output_img (optional): Filename to save the output ROC curve plot. Default is roc_curve.png.

Example

python beta.py data/Training_alpha_and_beta_final.csv data/Testing_alpha_and_beta.csv --output_img beta_roc.png

python alpha-beta.py data/Training_alpha_and_beta_final.csv data/Testing_alpha_and_beta.csv --output_img alpha_beta_roc.png

How It Works

1. Data Loading: The scripts load both training and testing datasets, selecting specific columns for model input.
2. Data Transformation: A transformation function converts amino acid frequency data for class 0 and class 1, calculates frequency ratios, and applies a logarithmic transformation.
3. Model Training and Evaluation: Three classifiers (Linear SVC, Logistic Regression, Random Forest) are trained and cross-validated. ROC curves are plotted to compare the performance of each classifier.

Output

Each script generates an ROC curve plot saved as an image file. The plot includes the following details:

• Mean AUC (Area Under the Curve) across cross-validation folds.
• Test AUC for evaluating the final model on the test dataset.

---