EnDecon

EnDecon applies ensemble learning to deconvolve cell-type compositions from spatially resolved transcriptomics (SRT) spots to resolve cellular heterogeneity across tissue regions.

Key Features:

• Ensemble learning framework: Integrates multiple base deconvolution methods using a weighted optimization model that assigns weights based on method performance.
• Performance superiority: Simulation studies and comprehensive benchmarking demonstrate higher accuracy than individual deconvolution methods, with learned weights positively correlating with base-method performance.
• Real-dataset application: Identifies multiple cell types per spot, localizes cell types to specific spatial regions, and distinguishes patterns of spatial colocalization and enrichment in SRT datasets.
• Robustness to technical variation: Maintains consistent performance under variations in sequencing depth and spot size.
• Normalization considerations: Notes that most deconvolution methods perform optimally when data normalization aligns with the procedures described in their original publications.
• Benchmarking scale: Evaluated 14 deconvolution methods across four datasets as part of its performance assessment.

Scientific Applications:

• Tissue architecture mapping: Provides quantitative cell-type composition estimates for analysis of tissue architecture and function.
• Spatial heterogeneity and regionalization: Enables localization of cell types and identification of region-specific enrichment and heterogeneity.
• Analysis of spatial co-localization: Supports investigation of spatial colocalization patterns and potential cellular interactions within tissues.

Methodology:

Integrates multiple base deconvolution methods via a weighted optimization ensemble that learns weights from method performance, with validation by simulation studies and benchmarking across 14 deconvolution methods and four datasets and comparisons of normalization strategies.