DeepTraCDR: Prediction Cancer Drug Response using multimodal deep learning with Transformers

Method

Requirements

To run this project, you need to install the required dependencies first. Execute the following command in your terminal or command prompt:

pip install -r requirements.txt

DeepTraCDR Model Overview

DeepTraCDR is a modular model consisting of Common Modules and Experimental Modules.

Common Modules

• Data: Includes datasets for model training and evaluation:
  • GDSC: Contains cell_drug.csv (log IC50 matrix), cell_drug_binary.csv (binary matrix), merged_file_GDSC.csv (gene expression), drug_feature.csv (drug fingerprints), null_mask.csv (null values), and threshold.csv (sensitivity threshold).
  • CCLE: Similar to GDSC with cell_drug.csv, cell_drug_binary.csv, merged_file.csv(gene expression), and drug_feature.csv.
  • PDX: Includes pdx_response.csv (binary patient-drug matrix), pdx_exprs.csv (gene expression), pdx_null_mask.csv (null values), and drug_feature.csv.
  • TCGA: Contains patient_drug_binary.csv (binary matrix), tcga_exprs.csv (gene expression), tcga_null_mask.csv (null values), and drug_feature.csv.

Experimental Modules

The experimental modules are organized into the following directories, each containing a main.py script to run the respective experiment:

• case_study: Contains scripts for case study experiments (e.g., main_case_study.py).
• Scenario1: Includes experiments for random clearing cross-validation (Random) and regression (Regression).
• Scenario2: Includes experiments for single row/column clearing (new) and targeted drug experiments (Target).
• Scenario3: Includes external validation experiments from in vitro to in vivo (External).

Each main.py script outputs true and predicted test data values after multiple cross-validations. The utils.py file supports performance analysis with metrics like AUC, AUPRC, ACC, F1, and MCC. The model is built using PyTorch with CUDA support.