How pfizer-opensource/pfabnet-viscosity works
Main Components
A. Data Processing Pipeline
• Raw Data Input (antibody sequences and viscosity measurements)
• Homology Model Generation (using Bioluminate)
• Data Preprocessing Components
B. Model Architecture
• PfAbNet Core Model
• Training Pipeline
• Inference Pipeline
C. Analysis Components
• Attribution Analysis System
• Sensitivity Analysis System
• Validation System
Component Relationships
[Raw Data] -> [Preprocessing] -> [Model Training/Inference] -> [Analysis]
Specific Implementation Details
A. Data Layer (Bottom)
• Data files (CSV files in data/)
• Homology model directory
• External dependencies (Bioluminate, OpenEye toolkits)
B. Processing Layer (Middle)
• Preprocessing components (from notebooks/1_preprocess.ipynb)
• Model building components (from pfabnet/model.py)
• Training system (from pfabnet/trainer.py)
C. Analysis Layer (Top)
• Attribution generation (pfabnet/generate_attributions.py)
• Validation components
• Sensitivity analysis tools
Important Elements to Highlight
A. Two Main Model Variants:
• PfAbNet (pfabnet/ directory)
• PfAbNet-Eisenberg (pfabnet_eisenberg/ directory)
B. Pipeline Stages:
• Data preprocessing
• Model building
• Training
• Inference
• Analysis
Technical Details
Connections between:
• Dataset handling (dataset.py)
• Model architecture (model.py)
• Training process (trainer.py)
• Prediction pipeline (predict.py)
Workflow Representation
A workflow section showing the sequence: