MEFISTO

MEFISTO performs dimensionality reduction and pattern identification in multi-modal omics datasets by explicitly integrating known spatial and temporal relationships among samples.

Key Features:

• Spatio-Temporal Integration: Incorporates known spatial and temporal relationships between samples to model continuous structures defined by covariates such as position or time.
• Dimensionality Reduction and Pattern Disentanglement: Performs unsupervised factor analysis to reduce dimensionality while distinguishing smooth sources of variation that change along covariates from other independent variation.
• Interpolation/Extrapolation: Enables interpolation and extrapolation to predict patterns at unseen timepoints or locations using modeled spatio-temporal dependencies.
• Multi-Dataset Integration: Aligns underlying patterns of variation across multiple related datasets in a data-driven manner to integrate diverse omics types or longitudinal data.
• Multi-modal omics support: Targets analysis of high-dimensional, multi-modal omics datasets, including cases with spatially resolved measurements.
• Unsupervised approach: Operates without supervision to uncover latent factors and biological processes from observed data.

Scientific Applications:

• Evolutionary organ development atlases: Applied to an evolutionary atlas of mammalian organ development to reveal conserved developmental programs and programs that diverged evolutionarily.
• Longitudinal microbiome studies: Used in infant microbiome time series to identify birth mode and diet as contributors to microbiome heterogeneity over time.
• Spatially resolved transcriptomics: Analyzes gene expression data with known spatial coordinates to provide insights into tissue organization and function.

Methodology:

Uses factor analysis adapted to explicitly model spatio-temporal dependencies, separating smooth (covariate-driven) from non-smooth patterns of variation and enabling interpolation/extrapolation across covariate space.