High-Dimensional Multi-Study Robust Factor Model for Analyzing RNA Sequencing Data from Heterogeneous Sources

The amount of high-dimensional large-scale RNA sequencing data derived from multiple heterogeneous sources has increased exponentially in biological science. During data collection, significant technical noise or errors may occur. To robustly extract meaningful features from this type of data, we introduce a high-dimensional multi-study robust factor model, called MultiRFM, which learns latent features and accounts for the heterogeneity among sources. MultiRFM demonstrates significantly greater robustness compared to existing multi-study factor models and is capable of estimating study-specific factors that are overlooked by single-study robust factor models. Specifically,we utilize a multivariate t-distribution to model errors, capturing potential heavy tails, and incorporate both study-shared and study-specified factors to represent common and specific information among studies. For parameter estimation, we have designed a computationally efficient variational estimation approach. A step-wise singular value ratio method is proposed to determine the discrete tuning parameters. Extensive simulation studies indicate that MultiRFM surpasses state-of-the-art methods in terms of estimation accuracy across various scenarios. Real-world applications involving two RNA sequencing datasets demonstrate that MultiRFM outperforms competing methods in model fitting, prediction, and computational efficiency, significantly facilitating downstream tasks.