DeepSADR

=======

DeepSADR

DeepSADR: A Deep Learning Model Based on Subsequence Interaction and Adaptive Readout for Enhanced Drug Response Prediction in Cancer Patients

Overview

We propose DeepSADR, a transfer learning model for drug response prediction from cell lines to patients, built on subsequence interaction and adaptive readout. DeepSADR adopts a ’pre-training + fine-tuning’ strategy. We construct sub-sequence interaction graphs to explore the associations between drug and gene subsequences, thereby improving model performance and interpretability. To achieve effective model transfer in the drug response(domain), we introduce an adaptive readout function to learn domain-invariant drug response features, thereby improving the model’s predictive performance on patient data.

Data

'data/Cell : Includes cell lines geneomic profiles data, drug Smiles sequences, and drug response data.

'data/Patient': Includes patients geneomic profiles data, drug Smiles sequences, and drug response data.

'data /split_cell_lines.csv': Classification results for genes

Environment

You can create a conda environment for DeepSADR by ‘conda env create -f environment.yml‘.

Train and test

• step 1

  • pre-trian

    'python pretrain_model.py'
    

• step 2 (regression experiment):

  • fine-tune

    'python fine_tune_model.py'