PhEMD

PhEMD maps relationships among multicellular biospecimens profiled at single-cell resolution by computing Earth Mover's Distance (EMD) between sample-level cell-state distributions and constructing a low-dimensional "manifold of manifolds" embedding to reveal axes of phenotypic variation.

Key Features:

• Manifold-of-manifolds representation: Embeds each biospecimen as a point in a higher-level manifold where each point represents a collection of cells spanning a lower-level cell-state manifold.
• Earth Mover's Distance (EMD): Computes pairwise distances between samples using Earth Mover's Distance to compare sample-level cell-state distributions.
• Low-dimensional embedding: Constructs a low-dimensional embedding that reveals intrinsic manifold structures among biospecimens.
• Axes of phenotypic variation: Identifies axes of variation among complex biological specimens rather than focusing solely on individual cells.
• Cell subpopulational variation detection: Uncovers axes of cell subpopulational variation across diverse perturbation conditions.
• Application to drug-screen datasets: Demonstrated on a newly generated drug-screen dataset to resolve perturbation-driven phenotypic differences.
• Phenotype inference for unprofiled samples: Can infer phenotypes of biospecimens that were not directly profiled.
• Clinical patient-state mapping: Generates maps of patient-state space in clinical datasets and highlights sources of variation between patients.
• Compatibility with batch-effect correction: Designed to be compatible with leading batch-effect correction techniques for cross-study comparisons.
• Scalability and generalizability: Emphasizes scalability and applicability across different studies and datasets.
• Single-cell resolution focus: Specifically targets multicellular biospecimens profiled at single-cell resolution.

Scientific Applications:

• Comparative phenotyping: Compare phenotypic relationships among multicellular biospecimens to identify specimen-level variation.
• Perturbation analysis: Analyze drug-screen and other perturbation datasets to detect shifts in cell subpopulations and phenotypes.
• Phenotype prediction: Infer likely phenotypes for biospecimens that lack direct single-cell profiling.
• Clinical heterogeneity mapping: Map patient-state space to highlight heterogeneity and sources of variation across patients in clinical datasets.
• Cross-study integration: Enable comparisons of biospecimens across studies using batch-effect correction and the manifold-of-manifolds framework.

Methodology:

Computes pairwise sample distances using Earth Mover's Distance (EMD), represents samples as points in a "manifold of manifolds" where each point encodes a cell-state distribution, and constructs a low-dimensional embedding to reveal intrinsic manifold structures.