Federated generative imputation with variational and denoising autoencoders for pan-cancer genomics

In this article, the problem of missing data imputation in federated framework is studied using high dimensional Cancer Genome Atlas (TCGA) dataset. The imputation and prediction performances of generative autoencoder based algorithms VAE and DAE were tested on data with features missing completely at random (MCAR). The performance of the models was studied under different levels of heterogeneity and missing data in training. As baseline models, mean and KNN imputation were used, and the performances were evaluated using Mean Square Error (MSE) for imputation and accuracy for the prediction task. The biological relevance of the imputed datasets was also analyzed using three different techniques. The results indicate that the datasets imputed by VAE and DAE maintain the biological relevance of the datasets better in comparison to mean imputation with VAE being more resilient to a higher percentage of missing data in training during imputation, as compared to DAE.