ML_Final_Project: Protein Feature Extraction Pipeline

This project builds a clean and reproducible pipeline to extract PSSM and structural features from protein sequences, starting only from g_data.csv.

---

Step 1: FASTA File Preparation

• Input: g_data.csv inside the data/ directory.
• Goal: Generate individual .fasta files, one per protein.

Handling Invalid Amino Acids ('U' → 'X')

Problem:
The original g_data.csv contained the amino acid 'U' (Selenocysteine), which causes errors in:

• SPOT-1D
• PSI-BLAST
• PSSM matrix generation

Solution:

• Manually replaced all occurrences of 'U' with 'X' directly in g_data.csv.
• Then generated .fasta files.

---

Step 2: SwissProt BLAST Database Setup

• Downloaded SwissProt database (uniprot_sprot.fasta) from UniProt FTP.
• Unzipped the database.
• Created a BLAST+ database using:

makeblastdb -in uniprot_sprot.fasta -dbtype prot

---

Step 3: PSSM Matrix Generation (PSI-BLAST)

• Ran PSI-BLAST (psiblast) for each .fasta file inside data/fasta_files/.
• Used the SwissProt BLAST database prepared earlier.
• Settings used:
  • Number of iterations: 3
  • E-value cutoff: 0.001
• Saved the output .pssm files inside the pssm_outputs/ folder.

Command used to run multiple PSI-BLAST jobs in parallel:

bash code/run_psiblast.sh

---

Step 4: PSSM Feature Extraction

After generating .pssm files from PSI-BLAST, the following features were extracted:

Feature Set	Number of Features
PSSM-AAO (Amino Acid Occurrence)	20
PSSM-SD (Standard Deviation over Segments)	80
PSSM-SAC (Sequence Autocovariance)	100

Each .pssm file was processed using the following scripts:

• code/extract_pssm_aao.py → outputs features/pssm_aao.csv
• code/extract_pssm_sd.py → outputs features/pssm_sd.csv
• code/extract_pssm_sac.py → outputs features/pssm_sac.csv

Details:

• PSSM-AAO: Summation of scores across all amino acids.
• PSSM-SD: Mean and standard deviation values across 4 segments.
• PSSM-SAC: Autocovariance of scores across 5 segments with a lag of 10.

---

Step 5: SPOT-1D-Single Structural Prediction

• Cloned SPOT-1D-Single into the SPOT-1D-Single/ folder.
• Downloaded and extracted pre-trained models and standardization files (means_single.pkl, stds_single.pkl).
• Modified SPOT-1D-Single to:
  • Accept .fasta file paths directly.
  • Output predictions into the spot1d_outputs/ folder.

Notes:

• Always pass the output folder (spot1d_outputs/) as --save_path, not individual filenames.
• SPOT-1D automatically names each .csv file after the protein ID.

Batch Prediction Command:

python code/run_spot1d_batch.py

The batch script picks each .fasta file from data/fasta_files/, runs SPOT-1D prediction, and saves the corresponding .csv output in spot1d_outputs/.

---

Step 6: Structural Feature Extraction

Following the generation of SPOT-1D outputs, the following structural features were extracted:

Feature Set	Number of Features
ASA Bigram	1
SS3 Autocovariance (Lag 10)	30
Torsion Composition (Phi, Psi, Theta, Tau)	4

Step 7: Merging Features and Labels

• Cleaned Protein_IDs by removing > and whitespace.
• Merged all six features using Protein_ID as the key.
• Joined with labels from g_data.csv.

python merge_features.py
python merge_labels.py

Output: final_training_data.csv

---

Step 8: Model Training

Trained Models:

• SVM (RBF kernel) — C=3000, γ=0.005
• K-Nearest Neighbors
• Naive Bayes
• Random Forest
• Bagging Classifier
• ANN (MLPClassifier)

Evaluation:

• Train/Test Split (80/20)
• 10-Fold Cross-Validation
• Jackknife (LOOCV)

---

Step 9: Model Evaluation Results

Model	10-Fold CV Accuracy	Jackknife Accuracy
SVM	67.11%	67.88%
K-NN	68.25%	–
Naive Bayes	60.63%	–
Random Forest	73.05%	72.66%
Bagging	71.14%	71.89%
ANN (MLP)	72.26%	71.51%

✅ Best Performer: Random Forest
✅ Most Consistent: Random Forest & ANN

---

Step 10: Visualization

• 10-Fold CV accuracy comparison saved as:

Evaluation/accuracy_plot.png

---

Scripts used:

• code/extract_asa_bigram.py → outputs features/asa_bigram.csv
• code/extract_ss3_autocov.py → outputs features/ss3_autocov.csv
• code/extract_torsion_composition.py → outputs features/torsion_composition.csv

Details:

• ASA Bigram: Mean of ASA[i] × ASA[i+2] across sequence.
• SS3 Autocovariance: Measures autocorrelation for Coil (C), Helix (H), and Strand (E) probabilities across 10 lags.
• Torsion Composition: A simple means of torsion angles is Phi, Psi, Theta, Tau.

---

Final Folder Structure After Feature Extraction

ML_Final_project/
|
├── data/
│   └── fasta_files/          # All input .fasta files
├── pssm_outputs/              # PSI-BLAST output .pssm files
├── spot1d_outputs/            # SPOT-1D-Single output .csv files
├── features/                  # Extracted feature CSVs
│    ├── pssm_aao.csv
│    ├── pssm_sac.csv
│    ├── pssm_sd.csv
│    ├── asa_bigram.csv
│    ├── ss3_autocov.csv
│    └── torsion_composition.csv
├── code/
│    ├── run_psiblast.sh
│    ├── run_spot1d_batch.py
│    ├── extract_pssm_aao.py
│    ├── extract_pssm_sd.py
│    ├── extract_pssm_sac.py
│    ├── extract_asa_bigram.py
│    ├── extract_ss3_autocov.py
│    └── extract_torsion_composition.py
├── SPOT-1D-Single/             # SPOT-1D model code and models
├── README.md
├── environment.yml
└── requirements.txt

python generate_fasta.py
bash code/run_psiblast.sh
python code/run_spot1d_batch.py
python extract_all_features.py
python merge_features.py
python merge_labels.py
python train_svm.py
python train_cv.py
python train_jackknife.py

---

Key Implementation Details and Fixes

• Replaced invalid amino acid 'U' with 'X' early to prevent PSI-BLAST and SPOT-1D failures.
• Standardized .fasta file generation.
• Corrected file path handling inside SPOT-1D-Single.
• Created clean, parallelized batch processing.
• Safely handled missing output files and sequences during feature extraction.
• Cleaned inconsistent header formats
• Merged features and labels robustly
• Handled class imbalance (Fold2 had very few samples)

---