Differential Private Deep Learning Models for Analyzing Breast Cancer Omics Data.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• High-dimensional human genomic data is crucial for understanding diseases and drug responses.
• Genomic data contains sensitive private information, hindering public sharing and analysis.
• Deep learning (DL) models excel at analyzing genomic data but can leak training sample information.

Purpose of the Study:

• To develop a privacy-preserving framework for analyzing high-dimensional human genomic data.
• To enable accurate prediction of breast cancer status (BCS), cancer type (CT), and drug sensitivity using genomic data.
• To address the privacy concerns associated with sharing and analyzing sensitive genomic datasets.

Main Methods:

• Proposed a differential privacy-based deep learning (DP-DL) framework.
• Developed a differential private deep autoencoder (dpAE) for low-dimensional genomic data representation.
• Utilized dpAE features to build DP binary classifiers for BCS and CT prediction.
• Applied the framework to the Genomics of Drug Sensitivity in Cancer (GDSC) dataset for drug sensitivity prediction.

Main Results:

• The DP-DL framework achieved improved prediction performance for BCS and CT classification compared to previous DP methods.
• The framework demonstrated effective drug sensitivity prediction for 265 drugs using GDSC data.
• The proposed method successfully balances predictive accuracy with robust individual privacy protection.

Conclusions:

• The developed DP-DL framework offers a viable solution for privacy-preserving analysis of sensitive genomic data.
• This approach facilitates the advancement of biological insights from genomic data without compromising individual privacy.
• The framework shows promise for applications in precision medicine and personalized drug development.