How jimmyjbling/sgc-del-ml-wdr91 works
Project type and purpose
• This is a command-line ML screening pipeline for small-molecule SMILES.
• It provides two main functions:
• train.py – builds fingerprint-based ensemble models and pickles them under models/
• nominate.py (screening) – loads those models to predict binder probability (“PRED”) and confidence (“CONF”) on new SMILES inputs, writing results into nominations/
File structure
• Top-level directories:
• models/ – serialized ensemble models (pickles) named by fingerprint sets and ensemble size
• nominations/ – output CSV and .PREDS files for screening results
• scripts/ – two entry-point scripts: train.py (model training) and nominate.py (screening)
• readme.md – high-level description of model use and confidence metric
• requirements.txt – Python dependencies (e.g., scikit-learn, RDKit)
README for additional insights
• Describes two screening workflows: initial (beCFP6 ensemble) and rescreening (combined fingerprints)
• Explains “PRED” vs. “CONF” and why both matter (entropy-based confidence)
• Notes that train.py also contains fingerprinting functions (chemical featurization via RDKit)
Components
a. Main components:
• Data Input (SMILES files or lists)
• Fingerprint Generator (module within train.py & nominate.py)
• Model Training (train.py → ensemble trainer → models/*.pkl)
• Model Storage (models directory)
• Screening Engine (nominate.py / screen_smiles function)
• Output Writer (nominations/*.csv and .PREDS)
• External Dependencies (RDKit for fingerprints, scikit-learn for modeling)
b. Relationships & interactions:
• Data Input → Fingerprint Generator → Model Training → Model Storage
• Data Input → Fingerprint Generator → Screening Engine (loads models) → Output Writer
• Both scripts read requirements.txt (implicit) and import shared fingerprinting code
c. Architectural patterns / design principles:
• Modular separation: training vs. inference
• File-based persistence (pickles and CSV outputs)
• Single-process CLI tools (no microservices or databases)
• Reuse of fingerprinting code across train.py & nominate.py
d. Key technologies / libraries:
• Python 3.x
• RDKit – chemical fingerprint generation
• scikit-learn – ensemble classifiers & confidence (entropy)
• Pickle – model serialization
• CSV/.PREDS – output formats