Final project for CS549 Machine Learning course.
Multi-label classification of protein structures using features from the Protein Data Bank (PDB). Given a protein's biophysical properties (molecular weight, residue count, pH, solvent density, etc.), this pipeline predicts functional class membership across 23 categories, including hydrolase, transferase, DNA/RNA binding, and immune-related proteins.
PDB classification labels are stored as free-text strings (e.g., "TRANSFERASE/HYDROLASE/DNA-BINDING PROTEIN"), which makes them difficult to use directly as training labels for machine learning. A single protein can belong to multiple functional categories, so a standard single-label encoding doesn't capture the biology.
This project addresses both problems:
Text parsing into structured labels: A keyword-matching engine converts free-text classification strings into a 23-class binary target matrix, handling ambiguous cases (e.g., distinguishing "DNA binding" from generic "binding") with explicit priority logic.
Multi-class classification: Three models (Random Forest, Decision Tree, Neural Network) are trained and evaluated on the resulting labeled dataset.
PDB CSV data
│
▼
[ multiclass_matrix.py ]
Parse free-text classification labels
Map keywords → 23 functional subclasses
Handle special cases (binding disambiguation, DNA/RNA logic)
Output: binary target matrix + biophysical features
│
▼
[ dataprocessing.py ]
Clean missing values
Encode categorical features (LabelEncoder)
Train/test split
│
├──→ Random Forest (scikit-learn)
├──→ Decision Tree (scikit-learn)
└──→ Neural Network (TensorFlow/Keras)
│
▼
[ Evaluation ]
Accuracy, Precision, Recall, F1 (weighted)
The target matrix encodes membership across these 23 categories, derived from keyword frequency analysis of PDB classification text (threshold: >1,000 occurrences in the dataset):
| Category | Category | Category |
|---|---|---|
| hydrolase | transferase | oxidoreductase |
| DNA_RNA_binding | protein_binding | other_binding |
| inhibitor | transport | DNA |
| RNA | transcription | immune |
| structural | isomerase | signal |
| ligase | viral | genomics |
| metal | membrane | chaperone |
| adhesion | regulation |
Binding disambiguation logic ensures that, for example, a protein labeled "DNA-BINDING" is assigned to DNA_RNA_binding rather than other_binding, and that the DNA flag is not redundantly set when DNA_RNA_binding already captures the relationship.
- Python 3.10+
- pandas
- numpy
- scikit-learn
- tensorflow
Install dependencies:
pip install pandas numpy scikit-learn tensorflowThis pipeline expects a CSV export from the Protein Data Bank containing at minimum these columns:
structureId, classification, macromoleculeType, residueCount, resolution, structureMolecularWeight, densityMatthews, densityPercentSol, phValue
PDB data can be downloaded from the RCSB PDB search interface.
Step 1: Generate the multi-class target matrix
importpandasaspdfrommulticlass_matriximportcreate_multiclass_matrixdata=pd.read_csv('pdb_data_no_dups.csv')
labeled_data=create_multiclass_matrix(data, matrix_creation_log='False')This outputs a CSV (matrix_data_set.csv) with binary subclass columns joined to the original biophysical features. Enable logging with matrix_creation_log='True' to inspect keyword matching decisions row by row.
Step 2: Run classification
python main.pySelect a CSV file when prompted, then choose a classifier (Random Forest, Decision Tree, or Neural Network) from the menu.
protein-structure-classification/
├── main.py # Entry point: file selection and model menu
├── multiclass_matrix.py # Free-text → binary target matrix conversion
├── dataprocessing.py # Feature preparation and model training/evaluation
├── filemanager.py # CSV file loading via file dialog
└── README.md
Python, pandas, NumPy, scikit-learn (RandomForestClassifier, DecisionTreeClassifier, LabelEncoder, train_test_split, classification metrics), TensorFlow/Keras (Sequential, Dense, Adam)
Reina Hastings - GitHub