Classification of contacts in protein structures

This software uses machine learning algorithms to classify residue-residue contacts within protein structures into categories as defined by the RING software (HBOND, VDW, SBOND, etc.) using structural and physicochemical features.

Project structure

BioinformaticsProject-main/
├── 3di_model/             # FoldSeek model files
├── config/                # Configuration files
├── data/                  # Input data and PDB structures
├── models/                # Trained models
├── output_3di/            # 3Di encoding output
├── output_features/       # Structural feature output
├── report/                # Figures and plots for analysis
├── scripts/               # Main scripts for feature extraction, training, evaluation
├── requirements.txt       # Python dependencies
└── README.md              # Project documentation

Description of features and data

Features used

Each contact is described using:

s_ss8 : Secondary structure from DSSP
s_rsa : Relative Solvent Accessibility
s_phi : Phi torsion angle
s_psi : Psi torsion angle
s_a1 : Atchley factor 1 - Hydrophilicity vs. hydrophobicity s_a2 : Atchley factor 2 - Secondary structure propensity
s_a3 : Atchley factor 3 - Molecular size s_a4 : Atchley factor 4 - Codon usage / Polarizability s_a5 : Atchley factor 5 - Electrostatic charge
s_3di_state : 3Di FoldSeek numeric state ID
s_3di_letter : 3Di FoldSeek letter

These are computed for both the source and target residues.

Prediction classes

The following are the supported contact types, matching RING nomenclature:

• Hydrogen bond (HBOND)
• Van der Waals (VDW)
• Disulfide bridge (SBOND)
• Salt bridge (IONIC)
• π–π stacking (PIPISTACK)
• π–cation (PICATION)
• π–hydrogen bond (PIHBOND)
• Metal ion coordination (METAL_ION)
• Halogen bond (HALOGEN)
• Unclassified

Instructions for software setup and usage

1. Install dependencies and DSSP

To install all packages used: pip install -r requirements.txt

2. Configuration

All paths and parameters are stored in config/config.json. Update this file to match your system setup.

3. Execution

a. Models and model retraining

Model training is performed in the following Jupyter notebook: train_model.ipynb

This script performs the following steps:

1. Loads training data from the path specified in config/config.json
2. Preprocesses the features
3. Visualizes feature distributions and correlations
4. Trains and evaluates multiple models, including:
   • Naive Bayes
   • Random Forest
   • LightGBM
5. Saves trained models as .pkl files into the models/ directory
6. Generates evaluation metrics such as:
   • Classification report
   • Confusion matrix
   • Matthews Correlation Coefficient (MCC)
   • Balanced Accuracy
   • Average Precision Score
   • Area Under ROC Curve (AUC)

To execute the full training and evaluation process, run this command in your terminal:

jupyter notebook train_model.ipynb

b. Model application to new .pdb structure

Prediction is performed using the notebook: model_prediction.ipynb

This script performs the following steps:

1. Extracts structural and 3Di features from the input PDB file using:
   • scripts/calc_features.py
   • scripts/calc_3di.py (within run_feature_extraction from scripts/runFeatureExtraction.py)
2. Visualizes extracted features using:
   • scripts/data_visualization.py
3. Loads a pre-trained model from path defined in config/config.json (e.g., Random Forest or Naive Bayes model in models/ directory)
4. Applies the model to the newly extracted contact features
5. Outputs predicted contact types and their classification scores

To apply a model to a new protein structure run the notebook using:

jupyter notebook model_prediction.ipynb

Ensure that:

• The input file exists in the data/pdbs/ directory
• Paths in config/config.json are set correctly
• Feature extraction tools (DSSP, FoldSeek) are installed and configured

Authors

Christina Caporale Natalya Lavrenchuk