How soufanom/sgcl-dpi works
Project Type and Purpose
• SGCL-DPI is a research-grade machine-learning pipeline for predicting drug–protein interactions. It implements a multi-stage workflow: data ingestion → feature and graph generation → model training (GNN + curriculum learning) → evaluation.
File-Structure Analysis
• Top-Level Directories:
• Data/ – raw and intermediate datasets (BindingDB, similarity graphs)
• simgraphmaker/ – graph and feature generation scripts
• proteins/ – protein‐feature extraction
• chemprotlinker/ – molecule–protein linking utilities
• gcns/ – main training script (sgcl_dpi.py)
• helpers/ – auxiliary parsers (stitch_data_parser.py)
• Configuration: requirements.txt governs dependencies; no deployment manifests (research prototype).
Components
a. Data Sources
• BindingDB-processed data (Data/BindingDB-processed)
• Any external reference data loaded in helpers/ or proteins/
b. Preprocessing Modules
• helpers/stitch_data_parser.py: parses and joins raw BindingDB entries
• proteins/protfeaturegen.py: computes protein descriptors
• chemprotlinker/chemprotlinker.py: links chemical and protein IDs if needed
c. Graph & Feature Generation (simgraphmaker/)
• simgraphmaker.py → features.pkl
• feature_similarity_computer.py → all‐pair similarity scores
• similarity_threshold_processor.py → thresholded similarity CSVs (drug & protein graphs)
Outputs: drug_similarity_sample.csv & protein_similarity_sample.csv in Data/SimilarityGraphs
d. Model Training (gcns/)
• sgcl_dpi.py orchestrates:
– Load features.pkl & similarity graphs
– Construct GNNs for drug & protein graphs
– Apply curriculum strategy (structure-guided sampling)
– Train, cross-validate, compute F1, AUC, AUPR
– Serialize metrics and (optionally) model checkpoints
Architectural Patterns & Design Principles
• Pipeline (linear DAG) with clear stage separation
• Modular scripts grouped by concern (data prep / graph prep / training)
• Data‐centric: each stage reads/writes artifacts (pkl, CSV)
• Reproducibility: versioned scripts, environment pinned by requirements.txt
Relationships & Data Flow
1. Data Sources → Preprocessing
2. Preprocessing OUTPUT → Graph & Feature Generation
3. Graph & Feature Generation OUTPUT → Model Training
4. Model Training OUTPUT → Evaluation reports & saved artifacts