MSDRP: a deep learning model based on multisource data for predicting drug response.

Area of Science:

• Computational biology
• Pharmacogenomics
• Machine learning in oncology

Background:

• Cancer heterogeneity complicates therapeutic outcomes, necessitating accurate in vitro drug response prediction for personalized medicine.
• Existing computational models often overlook crucial relationships between drugs and biological entities (targets, diseases, side effects) and pairwise drug-cell line interactions.

Purpose of the Study:

• To develop a novel deep learning framework, MSDRP, for predicting drug response in vitro.
• To enhance drug response prediction by integrating multi-source drug-biological entity associations and drug-cell line interactions.

Main Methods:

• Proposed MSDRP, a deep learning framework incorporating an interaction module for drug-cell line relationships.
• Utilized similarity network fusion algorithms to integrate multiple drug-biological entity associations.
• Employed feature vectors derived from multi-source drug similarity matrices.

Main Results:

• MSDRP outperformed state-of-the-art models in all performance measures across experiments.
• De novo and independent tests demonstrated MSDRP's excellent performance for predicting responses to new drugs.
• Case studies confirmed the model's interpretability and the utility of multi-source drug similarity features.

Conclusions:

• The MSDRP framework effectively predicts in vitro drug response by capturing complex interactions.
• Integrating multi-source drug-biological entity data and drug-cell line interactions significantly improves prediction accuracy.
• MSDRP offers a promising tool for advancing personalized cancer therapy through accurate drug response prediction.