Improving drug response prediction via integrating gene relationships with deep learning

Area of Science:

• Computational Biology
• Genomics
• Pharmacogenomics

Background:

• Predicting cancer drug response is vital for personalized medicine but hindered by tumor heterogeneity.
• Existing methods struggle with accuracy and robustness due to complex biological data.

Purpose of the Study:

• To develop a deep learning framework (DIPK) for accurate and robust prediction of cancer drug response.
• To integrate diverse biological data, including gene interactions, expression profiles, and molecular topologies, using self-supervised techniques.
• To evaluate DIPK's performance on known and novel cell lines/drugs and its applicability to single-cell and clinical data.

Main Methods:

• Developed Deep neural network Integrating Prior Knowledge (DIPK), a deep learning framework.
• Employed self-supervised learning to integrate gene interaction networks, gene expression data, and molecular topologies.
• Validated DIPK against existing methods using cancer cell line datasets and applied it to single-cell RNA sequencing and clinical data.

Main Results:

• DIPK demonstrated superior performance over existing methods in predicting drug response for both known and novel cell lines and drugs.
• The framework successfully extended its application to single-cell RNA sequencing data for response prediction and cell identification.
• DIPK accurately predicted higher paclitaxel response in the pathological complete response (pCR) group compared to the residual disease group in clinical data.

Conclusions:

• Integrating gene interaction relationships significantly improves drug response prediction accuracy.
• DIPK offers a robust and versatile tool for personalized cancer treatment by enhancing prediction capabilities.
• DIPK has the potential to aid clinical decision-making for individualized cancer therapy strategies.