Federerad hierarkisk klustring för RNA-sekvenseringsdata

Genetic data, and in particular RNA sequencing data analysis is crucial in our understanding of cancer, helping us prevent, diagnose and treat it. However, it is usually not easily accessible; because of its high sensitivity with regards to patient confidentiality, medical centers cannot often share it with research institutions. As a consequence, researchers are not able to leverage the power of big data in their models. Federated computations have been introduced to find a trade-off between privacy and model quality. Federated algorithms are executed over distributed datasets, without the medical centers ever sharing their raw data directly, by communicating locally aggregated quantities instead. While this is rather straightforward for gradient-based methods, it is more challenging to apply to many unsupervised algorithms, especially those involving pairwise distances between points located in different data centers. Yet, genetic data analysis pipelines – and notably the quality control or exploratory steps – rely on hierarchical agglomerative clustering, which crucially depends on such distances. We develop two approaches for federated agglomerative clustering of genetic data, and investigate their privacy properties, performance and complexity. We find the best quality for single and average linkage when testing them on bulk RNA sequencing data. The first approach closely follows the classic agglomerative procedure; it approximates local point groups as their centroid once they reach a critical size in order to share them. We empirically observe that the critical size can be as much as 10% of the total data size while maintaining a great performance. The second approach is based on computing the distance matrix centrally using random projections, a technique that can be employed for many other algorithms. It experimentally leads to satisfactory results for a significant dimensionality reduction. Privacy-wise, we attempt several attack models against this approach, which do not succeed on genetic data, indicating that collusion is the biggest threat for this method.