Representation transfer for differentially private drug sensitivity prediction

Abstract

Motivation Human genomic datasets often contain sensitive information that limits use and sharing of the data. In particular, simple anonymization strategies fail to provide sufficient level of protection for genomic data, because the data are inherently identifiable. Differentially private machine learning can help by guaranteeing that the published results do not leak too much information about any individual data point. Recent research has reached promising results on differentially private drug sensitivity prediction using gene expression data. Differentially private learning with genomic data is challenging because it is more difficult to guarantee privacy in high dimensions. Dimensionality reduction can help, but if the dimension reduction mapping is learned from the data, then it needs to be differentially private too, which can carry a significant privacy cost. Furthermore, the selection of any hyperparameters (such as the target dimensionality) needs to also avoid leaking private information. Results We study an approach that uses a large public dataset of similar type to learn a compact representation for differentially private learning. We compare three representation learning methods: variational autoencoders, principal component analysis and random projection. We solve two machine learning tasks on gene expression of cancer cell lines: cancer type classification, and drug sensitivity prediction. The experiments demonstrate significant benefit from all representation learning methods with variational autoencoders providing the most accurate predictions most often. Our results significantly improve over previous state-of-the-art in accuracy of differentially private drug sensitivity prediction. Availability and implementation Code used in the experiments is available at https://github.com/DPBayes/dp-representation-transfer.

Differential Private Deep Learning Models for Analyzing Breast Cancer Omics Data

“…Third, similar to (25), we collected pre-processed version of TCGA from https://xenabrowser.net/datapages/. We assembled the preprocessed version of the pan-cancer RNA-seq gene expression data from the TCGA while removing low expression genes.…”

“…However, Figures 2A, B shows that our proposed approach for the METABRIC dataset (for ER +/-classification) achieved improved accuracy and AUC for each of the predefined ∈s than the baseline. However, we can not use Niinimäki et al (25) approach for the METABRIC dataset as their approach requires a public dataset for representation learning, and METABRIC contains real private data.…”

“…Finally, we used these 2000-size DP representations of PD 1 patients to build our final dpClassM model. For the TCGA dataset, Figures 3A, B shows the comparison (mean accuracy and AUC of 10-fold cross-validation) of dpClassM against the baselines (25,31,32) with the same privacy budget (∈ = 1.0). In Figures 3A, B, the x-axis represents the ER status (ER+/-) and eight pairs of cancer types (number cases of the Supplementary Table S1) that we choose to perform our experiments.…”

“…Niinimäki et al (25) overcame the limitation of Honkela et al ( 8) by using the transfer learning in a differential privacy framework, which achieved state-of-the-art prediction performance (∈ = 10) using gene expression-based genomic datasets. They collected the TCGA and the GDSC datasets as a public and private datasets respectively.…”

“…It is possible to find a similar dataset from a private source. Besides, the amount of noise Niinimäki et al (25) insert into the data depends on the cardinality. Hence, this framework (25) can be used for a small training set only.…”