PDTSyn: A Parameter-Decomposed Transformer for Domain-Generalized Cell Line-Aware Drug Synergy Prediction

Overview

PDTSyn is a dual-channel transformer architecture for predicting drug synergy effects in cancer cell line contexts. It combines universal and cell-specific drug representations to achieve accurate drug combination synergy prediction.

Requirements

• torch
• dgl
• dgllife
• pandas
• numpy
• scikit-learn
• tensorboard
• rdkit-pypi

Usage

Basic Training

cd code
python run.py --dataset oneil --cv random --gpu 0

Complete Parameter Description

Data & Device Parameters

• --dataset (str, default: 'oneil'): Dataset type ('oneil', 'almanac')
• --gpu (int, default: 0): GPU device ID
• --cv (str, default: 'random'): Cross-validation type
  • random: Random splits
  • cell: Leave-one-cell-out validation
  • drug_pair: Leave drug pairs out
  • drug: Leave individual drugs out (ensures unseen drugs in test)

Model Hyperparameters

• --num_basis (int, default: 16): Number of basis functions for cell parameterization
• --hidden (int, default: 128): Main transformer dimension
• --heads (int, default: 4): Number of attention heads
• --layers (int, default: 6): Number of PDTransformer layers
• --mol_layers (int, default: 3): Number of molecular GCN layers
• --mol_hidden (int, default: 64): Molecular GCN hidden dimension

Training Hyperparameters

• --lr (float, default: 5e-4): Learning rate
• --epochs (int, default: 5000): Maximum training epochs
• --patience (int, default: 300): Early stopping patience (epochs)
• --folds (int, default: 10): Number of cross-validation folds
• --grad_clip (float, default: 5.0): Gradient clipping threshold

Loss Function Weights

• --kl_weight (float, default: 0.1): Weight for universal embedding KL regularization
• --marginal_weight (float, default: 0.1): Weight for cell-drug classification loss

Output Control

• --ckpt_name (str, default: None): Checkpoint file prefix
  • Default: ../ckpt/best_model_{fold}.ckpt
  • Custom: ../ckpt/{ckpt_name}_fold{fold}.ckpt
• --log_name (str, default: None): Log filename (overrides default)