Privacy-preserving cancer type prediction with homomorphic encryption

Esha Sarkar, Eduardo Chielle, Gamze Gürsoy, Yi Chen, Mark Gerstein, Michail ManiatakosScientific 2023 · DOI

TL;DR

This paper tackles the critical challenge of predicting cancer types from genomic data while preserving patient privacy. They achieve this by developing a novel feature engineering approach for genomic data and a highly optimized homomorphic encryption (HE) framework, resulting in a privacy-preserving model that is significantly faster than previous HE-based methods.

What Problem Does This Paper Address?

Precision medicine relies heavily on analyzing an individual's genetic information to tailor diagnoses and treatments. However, the sheer volume of genetic mutation data required to build accurate machine learning (ML) models for tasks like cancer type prediction presents a significant privacy hurdle. This data is often processed on powerful cloud servers, raising concerns about sensitive patient information being exposed. While homomorphic encryption (HE) offers a solution by allowing computations on encrypted data, its inherent computational overhead has historically made it impractical for complex, high-dimensional tasks like genomic analysis. This paper aims to bridge this gap by making privacy-preserving cancer prediction feasible and efficient.

Methods

The authors propose a multi-pronged approach:

  • Somatic Mutation Encoding: Instead of treating each individual mutation as a feature, they encode genes based on the impact and type of mutations they harbor. This involves a two-step process:
  • Feature Selection: Genes are selected based on mutation frequency and biological intuition, reducing the initial millions of mutations to a more manageable set of features.
  • Encoding Schemes: They explore various encoding methods to represent the presence and impact of SNVs (single nucleotide variations) and CNVs (copy number variations) in a way that's amenable to ML models and HE.
  • Statistical Feature Selection: After encoding, they employ statistical tests like the chi-square test to further refine the feature set, prioritizing those most informative for cancer type prediction.
  • Fast Matrix Multiplication for HE: A core contribution is a novel, optimized matrix multiplication algorithm specifically designed for HE. This algorithm is crucial for accelerating the linear operations (matrix multiplications and additions) that dominate ML inference, especially in the context of high-dimensional genomic data.
  • Privacy-Preserving Inference Protocol: They implement a complete protocol where patient data is encrypted by the client, processed homomorphically by the server, and the encrypted result is sent back for decryption. This ensures data remains encrypted throughout its lifecycle on the server.

Datasets

The study utilizes a large, real-world genomic dataset comprising over 2 million genetic mutations from 2713 patients across several cancer types. This dataset is derived from The Cancer Genome Atlas (TCGA) database. The authors note that this dataset is high-dimensional, with millions of potential features before their encoding and selection process.

Key Insights and Ideas

  • Biological Intuition is Key: Combining biological knowledge about mutation impact with statistical methods significantly improves feature selection and model performance, even for privacy-preserving methods.
  • Encoding Matters: How genomic mutations are represented as features has a profound impact on both prediction accuracy and the feasibility of HE computations.
  • Matrix Multiplication is the Bottleneck: For HE-based ML inference on high-dimensional data, the efficiency of matrix multiplication is paramount. Optimizing this operation is crucial for practical applications.
  • Latency is Critical for Real-World Use: Unlike some HE applications that prioritize throughput by batching, this work emphasizes low latency for individual patient predictions, which is more relevant for clinical decision-making.

Key Results

  • Improved Accuracy: The final privacy-preserving model achieved a micro-average Area Under the Curve (MAUC) of 0.98 and an accuracy of 83.61%, a significant improvement over previous baseline methods (70.08%).
  • Dramatic Speedup: The proposed HE-based matrix multiplication algorithm is approximately 550 times faster than standard matrix multiplication-based privacy-preserving models.
  • Efficient Inference: The system can predict cancer types for approximately 500 individuals in under a minute, with individual predictions taking around 1 second.
  • Dimensionality Reduction: Their feature engineering process reduced the initial millions of mutation features to around 43,000, making HE computation more tractable.
  • High Performance Binary Models: Individual binary models for each cancer type achieved test accuracies exceeding 90% and MAUCs above 0.97.

Limitations and Future Work

  • Honest-but-Curious Threat Model: The current work assumes an "honest-but-curious" (or semi-honest) threat model, where the cloud server performs computations correctly but might be curious about the data. More robust security models could be explored.
  • HE Overheads: While significantly improved, HE still introduces computational overheads compared to plaintext computation. Further optimizations for HE schemes and algorithms are always beneficial.
  • Generalizability: The authors acknowledge that healthcare models are difficult to port trivially across datasets. While their matrix multiplication algorithm is designed for general HE-based high-dimensional inference, further validation on diverse datasets would be valuable.
  • Future Work: The authors suggest exploring other HE schemes and further optimizing the matrix multiplication algorithm for even greater efficiency. They also highlight the potential for their approach to be generalized to other HE-based high-dimensional inference tasks beyond cancer prediction.

Summary updated

Abstract

Cancer genomics tailors diagnosis and treatment based on an individual's genetic information and is the crux of precision medicine. However, analysis and maintenance of high volume of genetic mutation data to build a machine learning (ML) model to predict the cancer type is a computationally expensive task and is often outsourced to powerful cloud servers, raising critical privacy concerns for patients' data. Homomorphic encryption (HE) enables computation on encrypted data, thus, providing cryptographic guarantees to protect privacy. But restrictive overheads of encrypted computation deter its usage. In this work, we explore the challenges of privacy preserving cancer type prediction using a dataset consisting of more than 2 million genetic mutations from 2713 patients for several cancer types by building a highly accurate ML model and then implementing its privacy preserving version in HE. Our solution for cancer type inference encodes somatic mutations based on their impact on the cancer genomes into the feature space and then uses statistical tests for feature selection. We propose a fast matrix multiplication algorithm for HE-based model. Our final model achieves 0.98 micro-average area under curve improving accuracy from 70.08 to 83.61% , being 550 times faster than the standard matrix multiplication-based privacy-preserving models. Our tool can be found at https://github.com/momalab/octal-candet .
BibTeX
@inproceedings{sarkar2023privacy,
  author = {Esha Sarkar and Eduardo Chielle and Gamze Gürsoy and Yi Chen and Mark Gerstein and Michail Maniatakos},
  title = {{Privacy-preserving cancer type prediction with homomorphic encryption}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-28481-8},
}
Figure