Efficient differentially private learning improves drug sensitivity prediction.

Area of Science:

• Computational biology
• Genomics
• Privacy-preserving machine learning

Background:

• Personalized recommendation systems require user data, posing privacy risks.
• Genomic data is crucial for precision medicine but highly sensitive and difficult to anonymize.
• Differential privacy offers a solution by ensuring individual patient data cannot be distinguished.

Purpose of the Study:

• To develop a robust differentially private regression method for accurate predictions from sensitive genomic data.
• To address the limitations of current differentially private learning methods in handling feasible data sizes and dimensionalities.

Main Methods:

• A new robust private regression method was developed.
• The method incorporates dimensionality reduction and outlier projection to minimize noise addition.
• The approach was evaluated on drug sensitivity prediction using genomic data.

Main Results:

• The proposed method achieved significant improvements in private drug sensitivity prediction accuracy.
• Predictive accuracy matched state-of-the-art non-private lasso regression with only 4x more samples.
• Effective performance was demonstrated even with moderately-sized datasets under strong differential privacy.

Conclusions:

• The differentially private regression method offers theoretical appeal, asymptotic efficiency, and good prediction accuracy.
• The method shows promise for practical applications in genomics and other fields requiring privacy-preserving analysis of sensitive data.