DeepCDR: a hybrid graph convolutional network for predicting cancer drug response

Source: DeepCDR: a hybrid graph convolutional network for predicting cancer drug response.

Overview

In this repository, I implemented DeepCDR, a multimodal network designed to predict cancer drug response from IC50 values, built from scratch.

Inference Results

Classification Performance	Value	Regression Performance	Value
Cross Entropy	3.9392	Root Mean Square Error	5.0984
Balanced Accuracy	0.8882	Pearson Correlation	0.9120

Installation

pip install -r requirements.txt

Train Configuration

output_dim: 100
dropout_prob: 0.1

mode: classification
seed: 42
epochs: 50
batch_size: 64
weight_decay: 1.0e-5
learning_rate: 1.0e-4
eta_min: 1.0e-6

Training

# Enter src directory
cd src/

# Train model
python train.py

Inference Configuration

batch_size: 20
mode: classification
identifier: 01-26-2025_15-hrs
weights: highest-balanced-accuracy.pth

Inference

# Enter src directory
cd src/

# Run Inference
python inference.py

Upon running inference, a table with the results for each sample will be saved in assets/inference-tables/{model_task}/{experiment_id}/results.csv.

Each table will contain the following fields per sample for downstream analysis:

• loss
• prediction
• target
• cell_line_id
• cancer_type
• drug_id