How xieresearchgroup/multidcp works
Main Components Organization
a) Input Processing Layer
• Chemical Structure Input (Graph Data)
• Gene Expression Profiles
• Chemical Substructure Data
• Dosage Information
b) Core Processing Modules
• Following key components:
• Graph Convolutional Network (GCN) Module
• Knowledge-enabled Autoencoder
• Chemical-Gene Interaction Module
• Dose Embedding Module
c) Training Pipeline
• Multi-task Learning Components:
• Chemical Transcriptomics Prediction
• Cell Viability Prediction
Component Relationships
• Feedback loops in the training process
• Integration points between different modules
Technical Implementation Details
Specific elements based on the file structure:
a) Model Components (from /models directory):
• Drug-gene attention mechanisms
• Graph degree convolution
• Multi-head attention
• Neural fingerprint processing
• -wise feedforward networks
b) Utility Layers (from /utils directory):
• Data processing pipelines
• Molecule utilities
• Metric calculations
Specific Implementation Notes
Annotations for:
• Pre-training phases (based on pretrain_multidcp.py)
• Fine-tuning steps (from ehill_multidcp_finetune.py)
• Inference pipelines (from drug_response_infer.py)
Technical Details
Labels indicating:
• Input data formats
• Model architecture specifications
• Training parameters
• Output formats