How volkamerlab/maxsmi works
Main Components
a) Data Processing Layer:
• SMILES Input Processing
• Augmentation Strategies Module
• Data Encoding Module
b) Model Layer:
• Deep Learning Models (CONV1D, CONV2D, RNN)
• Training Pipeline
• Early Stopping Module
• Ensemble Learning Component
c) Prediction/Evaluation Layer:
• Model Evaluation
• Prediction Pipeline
• Results Analysis
• Confidence Estimation
[Data Processing Layer] → [Model Layer] → [Prediction/Evaluation Layer]
[SMILES Input] → [Augmentation Strategies]
• No augmentation
• With duplication
• Without duplication
• Reduced duplication
• Estimated maximum
[Encoded Data] → [Neural Networks]
• CONV1D
• CONV2D
• RNN
Important Connections
• Data flow between augmentation strategies and model training
• Model training pipeline connection to evaluation
• Ensemble learning integration points
• Prediction pipeline for new compounds
Additional Elements
• Dataset handlers (ESOL, FreeSolv, lipophilicity, EGFR kinase)
• PyTorch utilities
• Model persistence (prediction_models directory)
• Results storage (output_ directory)
Technical Details to Highlight
• Integration with external libraries (DeepChem, PyTorch)
• Where configuration parameters are injected
• Evaluation metrics collection points
• Where model weights are saved/loaded
Specific Implementation Notes
• Training workflow (full_workflow.py)
• Early stopping workflow (full_workflow_earlystopping.py)
• Prediction workflow (prediction_unlabeled_data.py)
Use standardized symbols:
• Diamonds for decision points
• Rectangles for processes
This architecture should emphasize the project's unique aspects: SMILES augmentation strategies, multiple deep learning model support, and the confidence estimation feature of the Maxsmi approach.