Privacy-preserving cancer type prediction with homomorphic encryption

Area of Science:

• Computational biology
• Genomics
• Machine learning

Background:

• Cancer genomics is key to precision medicine, but handling large genetic datasets for ML models raises privacy concerns.
• Current methods often require outsourcing data analysis to cloud servers, increasing risks for patient data confidentiality.

Purpose of the Study:

• To develop a privacy-preserving machine learning model for accurate cancer type prediction.
• To address the computational expense and privacy challenges associated with analyzing large-scale cancer genomic data.

Main Methods:

• Utilized a dataset of over 2 million somatic mutations from 2713 cancer patients.
• Developed a machine learning model encoding mutation impact and employing statistical feature selection.
• Implemented a privacy-preserving version using homomorphic encryption (HE) with a novel fast matrix multiplication algorithm.

Main Results:

• Achieved a micro-average area under the curve of 0.98, significantly improving accuracy from 70.08% to 83.61%.
• The HE-based model demonstrated a 550-fold speed improvement over standard matrix multiplication methods.

Conclusions:

• The developed tool offers a computationally efficient and privacy-preserving solution for cancer type prediction.
• This approach enhances the feasibility of using sensitive genomic data for ML-driven cancer diagnostics.