Federated horizontally partitioned principal component analysis for biomedical applications

Abstract

Motivation Federated learning enables privacy preserving machine learning in the medical domain because the sensitive patient data remains with the owner and only parameters are exchanged between the data holders. The federated scenario introduces specific challenges related to the decentralized nature of the data, such as batch effects and differences in study population between the sites. Here, we investigate the challenges of moving classical analysis methods to the federated domain, specifically principal component analysis, a versatile and widely used tool, often serving as an initial step in machine learning and visualization workflows. We provide implementations of different federated PCA algorithms and evaluate them regarding their accuracy for high dimensional biological data using realistic sample distributions over multiple data sites, and their ability to preserve downstream analyses. Results Federated subspace iteration converges to the centralized solution even for unfavorable data distributions, while approximate methods introduce error. Larger sample sizes at the study sites lead to better accuracy of the approximate methods. Approximate methods may be sufficient for coarse data visualization, but are vulnerable to outliers and batch effects. Before the analysis, the PCA algorithm, as well as the number of eigenvectors should be considered carefully to avoid unnecessary communication overhead. Availability Simulation code and notebooks for federated PCA can be found at https://gitlab.com/roettgerlab/federatedPCA; the code for the federated app is available at https://github.com/AnneHartebrodt/fc-federated-pca Supplementary information Supplementary data are available at Bioinformatics Advances online.

Efficacy of federated learning on genomic data: a study on the UK Biobank and the 1000 Genomes Project

“…To mimic the situation where genetic testing companies from different parts of the world collaboratively predict ancestry, we split the 1000 Genomes Project data into 5 isolated nodes based on sample superpopulation (African, Native American, East Asian, European, Southern Asian), thus getting high internode heterogeneity. After a standard QC, we reduced dimensionality by applying federated PCA [40] to pruned SNPs and then trained local, federated and centralized multilayer perceptrons (MLPs) of identical architecture. The experiment setup is visualized in Figure 3 and described in more detail in the Methods section.…”

“…A fully federated solution requires all data to be prepared in a federated manner as well. In the case of ancestry prediction, dimensionality reduction is usually conducted via PCA, thus, we first pruned SNPs as displayed in Figure 3 to decrease computational load and then utilized federated PCA using the P-stack algorithm as described in [40]. The amount of communication used in federated PCA linearly depends on the number of input SNPs which affects the model accuracy.…”

“…We utilized the P-STACK method as described in [40], which involves sending a local eigenvalues vector and an eigenvector matrix from each client to the server. Further, the server stacks local PCA components and then performs a singular value decomposition (SVD) of the obtained joint matrix.…”

Unsupervised machine learning framework for discriminating major variants of concern during COVID-19

“…A high dimensional dataset is projected using PCA into an eigenspace that constitutes the direction of the largest variation illustrated by principal components. There are various drawbacks when using PCA, including the existence of abnormality that can lead to a recalculation of the PCA and result in unnecessary information disclosure [ 24 ]. Other than PCA, t-distributed stochastic neighbour embedding (t-SNE) [ 25 ] is widely used in the field of bioinformatics [ 26 ].…”

Unsupervised machine learning framework for discriminating major variants of concern during COVID-19

“…A high dimensional dataset are projected using PCA into an eigenspace that constitute the direction of largest variation illustrated by principal components. There are various drawbacks when using PCA including the existence of abnormality can lead to a recalculation of the PCA and result in unnecessary information disclosure [22]. Other than PCA, t-distributed stochastic neighbour embedding (t-SNE) [23] is widely used in the field of bioinformatics [24].…”