DeepDR: a deep learning library for drug response prediction.

Area of Science:

• Computational biology
• Pharmacogenomics
• Bioinformatics

Background:

• Accurate drug response prediction is crucial for advancing precision medicine and drug discovery.
• Deep learning (DL) shows promise in predicting drug response, but practical tools are lacking.
• Existing DL tools for drug response prediction are not user-friendly or comprehensive.

Purpose of the Study:

• Introduce DeepDR, the first deep learning library designed for drug response prediction.
• Simplify and automate the process of drug response modeling.
• Provide a flexible platform for building and evaluating various DL models.

Main Methods:

• DeepDR automates drug and cell featurization, model construction, training, and inference.
• The library supports three drug feature types and nine drug encoders.
• It also includes four cell feature types and nine cell encoders, plus two fusion modules.

Main Results:

• DeepDR enables the implementation of up to 135 distinct deep learning models.
• Benchmarking performance was explored using the DeepDR library.
• Optimal models identified through benchmarking are accessible via a visual interface.

Conclusions:

• DeepDR significantly lowers the barrier to entry for applying deep learning to drug response prediction.
• The library facilitates the development of more accurate predictive models for precision medicine.
• DeepDR supports researchers in accelerating drug discovery and personalized treatment strategies.