RNA-Binding Protein Prediction (PRBP-Style Random Forest)

This repository contains a small end-to-end pipeline to reproduce, at the molecular (protein) level, the RNA-binding protein (RBP) prediction setup described in:

Prediction of RNA-Binding Proteins Using a Random Forest Algorithm Combined with an RNA-Binding Residue Predictor (PRBP)

The implementation uses:

• EIPP: Evolution-Inspired Physicochemical Profiles derived from PSI-BLAST PSSMs
• AAC: Global amino-acid composition features
• Random Forests for binary protein-level RBP vs non-RBP classification

Contributors: Riddhi Mehta and Nirvisha Soni

---

Repository Structure

All files live under the Bioinformatics/ folder (GitHub will show this as the root of the repo):

Bioinformatics/
├── Positive_Final.fasta           # Final positive-class proteins (experimentally validated RBPs)
├── Negative_Final.fasta           # Final negative-class proteins (no known nucleic-acid binding)
├── all_sequences.fasta            # Combined positive + negative FASTA
│
├── split.py                       # Script: split all_sequences.fasta into one-FASTA-per-protein
├── generate_pssm.bat              # Windows batch script to run PSI-BLAST and generate PSSMs
│
├── mini_db.p*                     # Local PSI-BLAST (BLAST+) protein database files (mini_db)
│
├── split_seqs/                    # Folder of per-sequence FASTA files (created by split.py)
└── pssm_out/                      # Folder of per-sequence PSSMs (.pssm) from PSI-BLAST

The main modelling code is in:

Bioinformatics/PRBP_Prediction_EIPP_AAC_RF.ipynb

This Jupyter notebook:

1. Loads sequences and (optionally) PSSMs
2. Builds EIPP + AAC + length features
3. Trains a RandomForestClassifier
4. Evaluates with train/test split and 5-fold cross-validation
5. Prints metrics and example predictions for a few random proteins

---

Requirements

Python environment

• Python 3.8+
• Recommended: create a conda/venv environment

Python packages used in the notebook:

pip install numpy pandas scikit-learn matplotlib

BLAST+ / PSI-BLAST (for EIPP/PSSM features)

To recompute EIPP features from scratch you need:

• NCBI BLAST+ installed (so the psiblast command is available on your PATH)
• The provided mini_db.* BLAST database files in the same folder as the scripts

On Windows, you can install BLAST+ from the NCBI site and then add the bin folder to your PATH.

If you prefer not to re-run PSI-BLAST, you can:

• Use any pre-computed PSSMs already shipped in pssm_out/ (if present), or
• Skip the EIPP step and adapt the notebook to use only AAC features (for quick testing).

---

Getting Started

1. Clone or download the repository

   git clone https://github.com/riddhimehta15/Bioinformatics.git
   cd Bioinformatics/Bioinformatics

2. Create and activate a virtual environment (optional but recommended)

   Using conda:

   conda create -n rbp-prbp python=3.10
   conda activate rbp-prbp
   pip install numpy pandas scikit-learn matplotlib

   Or using venv:

   python -m venv .venv
   source .venv/bin/activate      # On Windows: .venv\Scripts\activate
   pip install numpy pandas scikit-learn matplotlib

3. Prepare per-sequence FASTA files

   The script split.py takes all_sequences.fasta and splits it into one FASTA file per protein in the split_seqs/ folder.

   From the Bioinformatics/ directory:

   python split.py

   After this step you should have many *.fasta files created under:

   Bioinformatics/split_seqs/
   

4. Generate PSSMs with PSI-BLAST

   This step computes a PSSM for each protein using the local mini_db BLAST database.

   From the Bioinformatics/ directory (or wherever generate_pssm.bat expects to be run; typically a pssm/ subfolder):

   generate_pssm.bat

   The script will:

   • Loop over split_seqs/*.fasta
   • Run psiblast for each sequence
   • Save an ASCII PSSM to pssm_out/<sequence_id>.pssm

   You can adjust parameters (e.g., number of iterations) inside generate_pssm.bat if needed.

5. Run the Random Forest notebook

   Start Jupyter:

   jupyter notebook

   Then open:

   Bioinformatics/PRBP_Prediction_EIPP_AAC_RF.ipynb
   

   and run all cells in order. The notebook will:

   • Parse the FASTA sequences (positive + negative)
   • Match them with available PSSMs in pssm_out/
   • Compute:
     • EIPP features from PSSMs (6 physicochemical groups × 20 amino acids = 120 dims)
     • AAC features (20-dimensional amino-acid composition)
     • Optional length feature
   • Train a Random Forest classifier with stratified train/test split
   • Report:
     • Accuracy, precision, recall, F1, ROC-AUC on the test set
     • 5-fold cross-validation scores
     • Top 20 most important features
     • Example predictions for 5 random proteins

---

Customization

• Changing train/test split
  Inside the notebook, adjust the test_size and RANDOM_STATE arguments in the train_test_split call.

• Changing Random Forest hyperparameters
  Modify the RandomForestClassifier parameters (e.g., n_estimators, max_depth, class_weight) to explore different models.

• Using only AAC features
  If computing PSSMs is too expensive, you can comment out the EIPP feature extraction and use only AAC + length features. This is useful for quick experimentation or for machines without BLAST+.

• Different datasets
  You can swap out Positive_Final.fasta and Negative_Final.fasta with your own datasets:

  • Make sure FASTA headers have unique IDs.
  • Rebuild all_sequences.fasta and re-run split.py + generate_pssm.bat.

---

Results

The notebook prints:

• Test set metrics (accuracy, precision, recall, F1, ROC-AUC)
• 5-fold cross-validation performance
• Feature importance plot for EIPP + AAC features
• Example predictions (predicted label and probability) for a few randomly chosen proteins.

Exact numbers will depend on:

• The dataset version
• PSSM generation parameters
• Random seed and train/test split

---

Contributors

• Riddhi Mehta
• Nirvisha Soni

---

Acknowledgements

This project is inspired by:

Wang, L., Huang, C., Yang, M.-Q., & Yang, J. Y. (2013). Prediction of RNA-binding proteins using a random forest algorithm combined with a novel RNA-binding residue predictor. Journal of Theoretical Biology, 1–9.

We also acknowledge the authors and maintainers of:

• UniProtKB/Swiss-Prot, for curated protein sequence data
• NCBI BLAST+, for PSI-BLAST and PSSM generation
• scikit-learn, NumPy, pandas, and matplotlib for model implementation and analysis