INSIDER

INSIDER performs interpretable sparse matrix decomposition via matrix factorization to decompose variation from multiple biological variables and their interactions in bulk RNA expression data for dimension reduction and adjusted expression profiling.

Key Features:

• Sparse matrix decomposition via matrix factorization: Performs interpretable sparse matrix decomposition using matrix factorization techniques for dimension reduction across high-dimensional datasets.
• Handling Multiple Biological Variables: Decomposes variation from multiple biological variables and their interactions into a shared low-rank latent space, unlike PCA or NMF which typically operate on two-dimensional sample × expression matrices.
• Elastic Net Penalty: Applies an elastic net penalty to induce sparsity while accounting for grouping effects of genes.
• Dimension Reduction (DR): Reduces dimensionality for datasets with three or more dimensions, extending beyond methods limited to two-dimensional data structures.
• Adjusted Expression Profiles: Computes adjusted expression profiles for specific biological variables while controlling for variation from other variables.
• Computational Efficiency and Missing Data Handling: Handles missing data and is computationally efficient, with performance comparable to or better than SDA in scenarios with complex missing data patterns typical of RNA-Seq datasets.
• Flexibility with Data Structures: Operates on data that cannot be organized into tensor format, avoiding a strict tensor requirement.

Scientific Applications:

• Clustering donors for disease subtyping: Clusters donors using RNA expression variation to support disease subtyping analyses.
• Revealing neurodevelopmental trajectories: Identifies developmental trajectories in neural tissues using datasets such as BrainSpan.
• Uncovering biological processes: Identifies biological processes that contribute to variables of interest, such as disease status and tissue type, and their interactions.

Methodology:

Uses matrix factorization for interpretable sparse matrix decomposition with an elastic net penalty to induce sparsity and account for gene grouping, constructs a shared low-rank latent space to decompose multiple biological variables and interactions, computes adjusted expression profiles, and handles missing RNA-Seq data; performance is compared to SDA.